1 // SPDX-License-Identifier: GPL-2.0-or-later
3 * raid5.c : Multiple Devices driver for Linux
4 * Copyright (C) 1996, 1997 Ingo Molnar, Miguel de Icaza, Gadi Oxman
5 * Copyright (C) 1999, 2000 Ingo Molnar
6 * Copyright (C) 2002, 2003 H. Peter Anvin
8 * RAID-4/5/6 management functions.
9 * Thanks to Penguin Computing for making the RAID-6 development possible
10 * by donating a test server!
16 * The sequencing for updating the bitmap reliably is a little
17 * subtle (and I got it wrong the first time) so it deserves some
20 * We group bitmap updates into batches. Each batch has a number.
21 * We may write out several batches at once, but that isn't very important.
22 * conf->seq_write is the number of the last batch successfully written.
23 * conf->seq_flush is the number of the last batch that was closed to
25 * When we discover that we will need to write to any block in a stripe
26 * (in add_stripe_bio) we update the in-memory bitmap and record in sh->bm_seq
27 * the number of the batch it will be in. This is seq_flush+1.
28 * When we are ready to do a write, if that batch hasn't been written yet,
29 * we plug the array and queue the stripe for later.
30 * When an unplug happens, we increment bm_flush, thus closing the current
32 * When we notice that bm_flush > bm_write, we write out all pending updates
33 * to the bitmap, and advance bm_write to where bm_flush was.
34 * This may occasionally write a bit out twice, but is sure never to
38 #include <linux/blkdev.h>
39 #include <linux/delay.h>
40 #include <linux/kthread.h>
41 #include <linux/raid/pq.h>
42 #include <linux/async_tx.h>
43 #include <linux/module.h>
44 #include <linux/async.h>
45 #include <linux/seq_file.h>
46 #include <linux/cpu.h>
47 #include <linux/slab.h>
48 #include <linux/ratelimit.h>
49 #include <linux/nodemask.h>
51 #include <trace/events/block.h>
52 #include <linux/list_sort.h>
57 #include "md-bitmap.h"
58 #include "raid5-log.h"
60 #define UNSUPPORTED_MDDEV_FLAGS (1L << MD_FAILFAST_SUPPORTED)
62 #define cpu_to_group(cpu) cpu_to_node(cpu)
63 #define ANY_GROUP NUMA_NO_NODE
65 #define RAID5_MAX_REQ_STRIPES 256
67 static bool devices_handle_discard_safely
= false;
68 module_param(devices_handle_discard_safely
, bool, 0644);
69 MODULE_PARM_DESC(devices_handle_discard_safely
,
70 "Set to Y if all devices in each array reliably return zeroes on reads from discarded regions");
71 static struct workqueue_struct
*raid5_wq
;
73 static inline struct hlist_head
*stripe_hash(struct r5conf
*conf
, sector_t sect
)
75 int hash
= (sect
>> RAID5_STRIPE_SHIFT(conf
)) & HASH_MASK
;
76 return &conf
->stripe_hashtbl
[hash
];
79 static inline int stripe_hash_locks_hash(struct r5conf
*conf
, sector_t sect
)
81 return (sect
>> RAID5_STRIPE_SHIFT(conf
)) & STRIPE_HASH_LOCKS_MASK
;
84 static inline void lock_device_hash_lock(struct r5conf
*conf
, int hash
)
85 __acquires(&conf
->device_lock
)
87 spin_lock_irq(conf
->hash_locks
+ hash
);
88 spin_lock(&conf
->device_lock
);
91 static inline void unlock_device_hash_lock(struct r5conf
*conf
, int hash
)
92 __releases(&conf
->device_lock
)
94 spin_unlock(&conf
->device_lock
);
95 spin_unlock_irq(conf
->hash_locks
+ hash
);
98 static inline void lock_all_device_hash_locks_irq(struct r5conf
*conf
)
99 __acquires(&conf
->device_lock
)
102 spin_lock_irq(conf
->hash_locks
);
103 for (i
= 1; i
< NR_STRIPE_HASH_LOCKS
; i
++)
104 spin_lock_nest_lock(conf
->hash_locks
+ i
, conf
->hash_locks
);
105 spin_lock(&conf
->device_lock
);
108 static inline void unlock_all_device_hash_locks_irq(struct r5conf
*conf
)
109 __releases(&conf
->device_lock
)
112 spin_unlock(&conf
->device_lock
);
113 for (i
= NR_STRIPE_HASH_LOCKS
- 1; i
; i
--)
114 spin_unlock(conf
->hash_locks
+ i
);
115 spin_unlock_irq(conf
->hash_locks
);
118 /* Find first data disk in a raid6 stripe */
119 static inline int raid6_d0(struct stripe_head
*sh
)
122 /* ddf always start from first device */
124 /* md starts just after Q block */
125 if (sh
->qd_idx
== sh
->disks
- 1)
128 return sh
->qd_idx
+ 1;
130 static inline int raid6_next_disk(int disk
, int raid_disks
)
133 return (disk
< raid_disks
) ? disk
: 0;
136 /* When walking through the disks in a raid5, starting at raid6_d0,
137 * We need to map each disk to a 'slot', where the data disks are slot
138 * 0 .. raid_disks-3, the parity disk is raid_disks-2 and the Q disk
139 * is raid_disks-1. This help does that mapping.
141 static int raid6_idx_to_slot(int idx
, struct stripe_head
*sh
,
142 int *count
, int syndrome_disks
)
148 if (idx
== sh
->pd_idx
)
149 return syndrome_disks
;
150 if (idx
== sh
->qd_idx
)
151 return syndrome_disks
+ 1;
157 static void print_raid5_conf (struct r5conf
*conf
);
159 static int stripe_operations_active(struct stripe_head
*sh
)
161 return sh
->check_state
|| sh
->reconstruct_state
||
162 test_bit(STRIPE_BIOFILL_RUN
, &sh
->state
) ||
163 test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
166 static bool stripe_is_lowprio(struct stripe_head
*sh
)
168 return (test_bit(STRIPE_R5C_FULL_STRIPE
, &sh
->state
) ||
169 test_bit(STRIPE_R5C_PARTIAL_STRIPE
, &sh
->state
)) &&
170 !test_bit(STRIPE_R5C_CACHING
, &sh
->state
);
173 static void raid5_wakeup_stripe_thread(struct stripe_head
*sh
)
174 __must_hold(&sh
->raid_conf
->device_lock
)
176 struct r5conf
*conf
= sh
->raid_conf
;
177 struct r5worker_group
*group
;
179 int i
, cpu
= sh
->cpu
;
181 if (!cpu_online(cpu
)) {
182 cpu
= cpumask_any(cpu_online_mask
);
186 if (list_empty(&sh
->lru
)) {
187 struct r5worker_group
*group
;
188 group
= conf
->worker_groups
+ cpu_to_group(cpu
);
189 if (stripe_is_lowprio(sh
))
190 list_add_tail(&sh
->lru
, &group
->loprio_list
);
192 list_add_tail(&sh
->lru
, &group
->handle_list
);
193 group
->stripes_cnt
++;
197 if (conf
->worker_cnt_per_group
== 0) {
198 md_wakeup_thread(conf
->mddev
->thread
);
202 group
= conf
->worker_groups
+ cpu_to_group(sh
->cpu
);
204 group
->workers
[0].working
= true;
205 /* at least one worker should run to avoid race */
206 queue_work_on(sh
->cpu
, raid5_wq
, &group
->workers
[0].work
);
208 thread_cnt
= group
->stripes_cnt
/ MAX_STRIPE_BATCH
- 1;
209 /* wakeup more workers */
210 for (i
= 1; i
< conf
->worker_cnt_per_group
&& thread_cnt
> 0; i
++) {
211 if (group
->workers
[i
].working
== false) {
212 group
->workers
[i
].working
= true;
213 queue_work_on(sh
->cpu
, raid5_wq
,
214 &group
->workers
[i
].work
);
220 static void do_release_stripe(struct r5conf
*conf
, struct stripe_head
*sh
,
221 struct list_head
*temp_inactive_list
)
222 __must_hold(&conf
->device_lock
)
225 int injournal
= 0; /* number of date pages with R5_InJournal */
227 BUG_ON(!list_empty(&sh
->lru
));
228 BUG_ON(atomic_read(&conf
->active_stripes
)==0);
230 if (r5c_is_writeback(conf
->log
))
231 for (i
= sh
->disks
; i
--; )
232 if (test_bit(R5_InJournal
, &sh
->dev
[i
].flags
))
235 * In the following cases, the stripe cannot be released to cached
236 * lists. Therefore, we make the stripe write out and set
238 * 1. when quiesce in r5c write back;
239 * 2. when resync is requested fot the stripe.
241 if (test_bit(STRIPE_SYNC_REQUESTED
, &sh
->state
) ||
242 (conf
->quiesce
&& r5c_is_writeback(conf
->log
) &&
243 !test_bit(STRIPE_HANDLE
, &sh
->state
) && injournal
!= 0)) {
244 if (test_bit(STRIPE_R5C_CACHING
, &sh
->state
))
245 r5c_make_stripe_write_out(sh
);
246 set_bit(STRIPE_HANDLE
, &sh
->state
);
249 if (test_bit(STRIPE_HANDLE
, &sh
->state
)) {
250 if (test_bit(STRIPE_DELAYED
, &sh
->state
) &&
251 !test_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
252 list_add_tail(&sh
->lru
, &conf
->delayed_list
);
253 else if (test_bit(STRIPE_BIT_DELAY
, &sh
->state
) &&
254 sh
->bm_seq
- conf
->seq_write
> 0)
255 list_add_tail(&sh
->lru
, &conf
->bitmap_list
);
257 clear_bit(STRIPE_DELAYED
, &sh
->state
);
258 clear_bit(STRIPE_BIT_DELAY
, &sh
->state
);
259 if (conf
->worker_cnt_per_group
== 0) {
260 if (stripe_is_lowprio(sh
))
261 list_add_tail(&sh
->lru
,
264 list_add_tail(&sh
->lru
,
267 raid5_wakeup_stripe_thread(sh
);
271 md_wakeup_thread(conf
->mddev
->thread
);
273 BUG_ON(stripe_operations_active(sh
));
274 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
275 if (atomic_dec_return(&conf
->preread_active_stripes
)
277 md_wakeup_thread(conf
->mddev
->thread
);
278 atomic_dec(&conf
->active_stripes
);
279 if (!test_bit(STRIPE_EXPANDING
, &sh
->state
)) {
280 if (!r5c_is_writeback(conf
->log
))
281 list_add_tail(&sh
->lru
, temp_inactive_list
);
283 WARN_ON(test_bit(R5_InJournal
, &sh
->dev
[sh
->pd_idx
].flags
));
285 list_add_tail(&sh
->lru
, temp_inactive_list
);
286 else if (injournal
== conf
->raid_disks
- conf
->max_degraded
) {
288 if (!test_and_set_bit(STRIPE_R5C_FULL_STRIPE
, &sh
->state
))
289 atomic_inc(&conf
->r5c_cached_full_stripes
);
290 if (test_and_clear_bit(STRIPE_R5C_PARTIAL_STRIPE
, &sh
->state
))
291 atomic_dec(&conf
->r5c_cached_partial_stripes
);
292 list_add_tail(&sh
->lru
, &conf
->r5c_full_stripe_list
);
293 r5c_check_cached_full_stripe(conf
);
296 * STRIPE_R5C_PARTIAL_STRIPE is set in
297 * r5c_try_caching_write(). No need to
300 list_add_tail(&sh
->lru
, &conf
->r5c_partial_stripe_list
);
306 static void __release_stripe(struct r5conf
*conf
, struct stripe_head
*sh
,
307 struct list_head
*temp_inactive_list
)
308 __must_hold(&conf
->device_lock
)
310 if (atomic_dec_and_test(&sh
->count
))
311 do_release_stripe(conf
, sh
, temp_inactive_list
);
315 * @hash could be NR_STRIPE_HASH_LOCKS, then we have a list of inactive_list
317 * Be careful: Only one task can add/delete stripes from temp_inactive_list at
318 * given time. Adding stripes only takes device lock, while deleting stripes
319 * only takes hash lock.
321 static void release_inactive_stripe_list(struct r5conf
*conf
,
322 struct list_head
*temp_inactive_list
,
326 bool do_wakeup
= false;
329 if (hash
== NR_STRIPE_HASH_LOCKS
) {
330 size
= NR_STRIPE_HASH_LOCKS
;
331 hash
= NR_STRIPE_HASH_LOCKS
- 1;
335 struct list_head
*list
= &temp_inactive_list
[size
- 1];
338 * We don't hold any lock here yet, raid5_get_active_stripe() might
339 * remove stripes from the list
341 if (!list_empty_careful(list
)) {
342 spin_lock_irqsave(conf
->hash_locks
+ hash
, flags
);
343 if (list_empty(conf
->inactive_list
+ hash
) &&
345 atomic_dec(&conf
->empty_inactive_list_nr
);
346 list_splice_tail_init(list
, conf
->inactive_list
+ hash
);
348 spin_unlock_irqrestore(conf
->hash_locks
+ hash
, flags
);
355 wake_up(&conf
->wait_for_stripe
);
356 if (atomic_read(&conf
->active_stripes
) == 0)
357 wake_up(&conf
->wait_for_quiescent
);
358 if (conf
->retry_read_aligned
)
359 md_wakeup_thread(conf
->mddev
->thread
);
363 static int release_stripe_list(struct r5conf
*conf
,
364 struct list_head
*temp_inactive_list
)
365 __must_hold(&conf
->device_lock
)
367 struct stripe_head
*sh
, *t
;
369 struct llist_node
*head
;
371 head
= llist_del_all(&conf
->released_stripes
);
372 head
= llist_reverse_order(head
);
373 llist_for_each_entry_safe(sh
, t
, head
, release_list
) {
376 /* sh could be readded after STRIPE_ON_RELEASE_LIST is cleard */
378 clear_bit(STRIPE_ON_RELEASE_LIST
, &sh
->state
);
380 * Don't worry the bit is set here, because if the bit is set
381 * again, the count is always > 1. This is true for
382 * STRIPE_ON_UNPLUG_LIST bit too.
384 hash
= sh
->hash_lock_index
;
385 __release_stripe(conf
, sh
, &temp_inactive_list
[hash
]);
392 void raid5_release_stripe(struct stripe_head
*sh
)
394 struct r5conf
*conf
= sh
->raid_conf
;
396 struct list_head list
;
400 /* Avoid release_list until the last reference.
402 if (atomic_add_unless(&sh
->count
, -1, 1))
405 if (unlikely(!conf
->mddev
->thread
) ||
406 test_and_set_bit(STRIPE_ON_RELEASE_LIST
, &sh
->state
))
408 wakeup
= llist_add(&sh
->release_list
, &conf
->released_stripes
);
410 md_wakeup_thread(conf
->mddev
->thread
);
413 /* we are ok here if STRIPE_ON_RELEASE_LIST is set or not */
414 if (atomic_dec_and_lock_irqsave(&sh
->count
, &conf
->device_lock
, flags
)) {
415 INIT_LIST_HEAD(&list
);
416 hash
= sh
->hash_lock_index
;
417 do_release_stripe(conf
, sh
, &list
);
418 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
419 release_inactive_stripe_list(conf
, &list
, hash
);
423 static inline void remove_hash(struct stripe_head
*sh
)
425 pr_debug("remove_hash(), stripe %llu\n",
426 (unsigned long long)sh
->sector
);
428 hlist_del_init(&sh
->hash
);
431 static inline void insert_hash(struct r5conf
*conf
, struct stripe_head
*sh
)
433 struct hlist_head
*hp
= stripe_hash(conf
, sh
->sector
);
435 pr_debug("insert_hash(), stripe %llu\n",
436 (unsigned long long)sh
->sector
);
438 hlist_add_head(&sh
->hash
, hp
);
441 /* find an idle stripe, make sure it is unhashed, and return it. */
442 static struct stripe_head
*get_free_stripe(struct r5conf
*conf
, int hash
)
444 struct stripe_head
*sh
= NULL
;
445 struct list_head
*first
;
447 if (list_empty(conf
->inactive_list
+ hash
))
449 first
= (conf
->inactive_list
+ hash
)->next
;
450 sh
= list_entry(first
, struct stripe_head
, lru
);
451 list_del_init(first
);
453 atomic_inc(&conf
->active_stripes
);
454 BUG_ON(hash
!= sh
->hash_lock_index
);
455 if (list_empty(conf
->inactive_list
+ hash
))
456 atomic_inc(&conf
->empty_inactive_list_nr
);
461 #if PAGE_SIZE != DEFAULT_STRIPE_SIZE
462 static void free_stripe_pages(struct stripe_head
*sh
)
467 /* Have not allocate page pool */
471 for (i
= 0; i
< sh
->nr_pages
; i
++) {
479 static int alloc_stripe_pages(struct stripe_head
*sh
, gfp_t gfp
)
484 for (i
= 0; i
< sh
->nr_pages
; i
++) {
485 /* The page have allocated. */
491 free_stripe_pages(sh
);
500 init_stripe_shared_pages(struct stripe_head
*sh
, struct r5conf
*conf
, int disks
)
507 /* Each of the sh->dev[i] need one conf->stripe_size */
508 cnt
= PAGE_SIZE
/ conf
->stripe_size
;
509 nr_pages
= (disks
+ cnt
- 1) / cnt
;
511 sh
->pages
= kcalloc(nr_pages
, sizeof(struct page
*), GFP_KERNEL
);
514 sh
->nr_pages
= nr_pages
;
515 sh
->stripes_per_page
= cnt
;
520 static void shrink_buffers(struct stripe_head
*sh
)
523 int num
= sh
->raid_conf
->pool_size
;
525 #if PAGE_SIZE == DEFAULT_STRIPE_SIZE
526 for (i
= 0; i
< num
; i
++) {
529 WARN_ON(sh
->dev
[i
].page
!= sh
->dev
[i
].orig_page
);
533 sh
->dev
[i
].page
= NULL
;
537 for (i
= 0; i
< num
; i
++)
538 sh
->dev
[i
].page
= NULL
;
539 free_stripe_pages(sh
); /* Free pages */
543 static int grow_buffers(struct stripe_head
*sh
, gfp_t gfp
)
546 int num
= sh
->raid_conf
->pool_size
;
548 #if PAGE_SIZE == DEFAULT_STRIPE_SIZE
549 for (i
= 0; i
< num
; i
++) {
552 if (!(page
= alloc_page(gfp
))) {
555 sh
->dev
[i
].page
= page
;
556 sh
->dev
[i
].orig_page
= page
;
557 sh
->dev
[i
].offset
= 0;
560 if (alloc_stripe_pages(sh
, gfp
))
563 for (i
= 0; i
< num
; i
++) {
564 sh
->dev
[i
].page
= raid5_get_dev_page(sh
, i
);
565 sh
->dev
[i
].orig_page
= sh
->dev
[i
].page
;
566 sh
->dev
[i
].offset
= raid5_get_page_offset(sh
, i
);
572 static void stripe_set_idx(sector_t stripe
, struct r5conf
*conf
, int previous
,
573 struct stripe_head
*sh
);
575 static void init_stripe(struct stripe_head
*sh
, sector_t sector
, int previous
)
577 struct r5conf
*conf
= sh
->raid_conf
;
580 BUG_ON(atomic_read(&sh
->count
) != 0);
581 BUG_ON(test_bit(STRIPE_HANDLE
, &sh
->state
));
582 BUG_ON(stripe_operations_active(sh
));
583 BUG_ON(sh
->batch_head
);
585 pr_debug("init_stripe called, stripe %llu\n",
586 (unsigned long long)sector
);
588 seq
= read_seqcount_begin(&conf
->gen_lock
);
589 sh
->generation
= conf
->generation
- previous
;
590 sh
->disks
= previous
? conf
->previous_raid_disks
: conf
->raid_disks
;
592 stripe_set_idx(sector
, conf
, previous
, sh
);
595 for (i
= sh
->disks
; i
--; ) {
596 struct r5dev
*dev
= &sh
->dev
[i
];
598 if (dev
->toread
|| dev
->read
|| dev
->towrite
|| dev
->written
||
599 test_bit(R5_LOCKED
, &dev
->flags
)) {
600 pr_err("sector=%llx i=%d %p %p %p %p %d\n",
601 (unsigned long long)sh
->sector
, i
, dev
->toread
,
602 dev
->read
, dev
->towrite
, dev
->written
,
603 test_bit(R5_LOCKED
, &dev
->flags
));
607 dev
->sector
= raid5_compute_blocknr(sh
, i
, previous
);
609 if (read_seqcount_retry(&conf
->gen_lock
, seq
))
611 sh
->overwrite_disks
= 0;
612 insert_hash(conf
, sh
);
613 sh
->cpu
= smp_processor_id();
614 set_bit(STRIPE_BATCH_READY
, &sh
->state
);
617 static struct stripe_head
*__find_stripe(struct r5conf
*conf
, sector_t sector
,
620 struct stripe_head
*sh
;
622 pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector
);
623 hlist_for_each_entry(sh
, stripe_hash(conf
, sector
), hash
)
624 if (sh
->sector
== sector
&& sh
->generation
== generation
)
626 pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector
);
630 static struct stripe_head
*find_get_stripe(struct r5conf
*conf
,
631 sector_t sector
, short generation
, int hash
)
633 int inc_empty_inactive_list_flag
;
634 struct stripe_head
*sh
;
636 sh
= __find_stripe(conf
, sector
, generation
);
640 if (atomic_inc_not_zero(&sh
->count
))
644 * Slow path. The reference count is zero which means the stripe must
645 * be on a list (sh->lru). Must remove the stripe from the list that
646 * references it with the device_lock held.
649 spin_lock(&conf
->device_lock
);
650 if (!atomic_read(&sh
->count
)) {
651 if (!test_bit(STRIPE_HANDLE
, &sh
->state
))
652 atomic_inc(&conf
->active_stripes
);
653 BUG_ON(list_empty(&sh
->lru
) &&
654 !test_bit(STRIPE_EXPANDING
, &sh
->state
));
655 inc_empty_inactive_list_flag
= 0;
656 if (!list_empty(conf
->inactive_list
+ hash
))
657 inc_empty_inactive_list_flag
= 1;
658 list_del_init(&sh
->lru
);
659 if (list_empty(conf
->inactive_list
+ hash
) &&
660 inc_empty_inactive_list_flag
)
661 atomic_inc(&conf
->empty_inactive_list_nr
);
663 sh
->group
->stripes_cnt
--;
667 atomic_inc(&sh
->count
);
668 spin_unlock(&conf
->device_lock
);
674 * Need to check if array has failed when deciding whether to:
676 * - remove non-faulty devices
679 * This determination is simple when no reshape is happening.
680 * However if there is a reshape, we need to carefully check
681 * both the before and after sections.
682 * This is because some failed devices may only affect one
683 * of the two sections, and some non-in_sync devices may
684 * be insync in the section most affected by failed devices.
686 * Most calls to this function hold &conf->device_lock. Calls
687 * in raid5_run() do not require the lock as no other threads
688 * have been started yet.
690 int raid5_calc_degraded(struct r5conf
*conf
)
692 int degraded
, degraded2
;
697 for (i
= 0; i
< conf
->previous_raid_disks
; i
++) {
698 struct md_rdev
*rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
699 if (rdev
&& test_bit(Faulty
, &rdev
->flags
))
700 rdev
= rcu_dereference(conf
->disks
[i
].replacement
);
701 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
))
703 else if (test_bit(In_sync
, &rdev
->flags
))
706 /* not in-sync or faulty.
707 * If the reshape increases the number of devices,
708 * this is being recovered by the reshape, so
709 * this 'previous' section is not in_sync.
710 * If the number of devices is being reduced however,
711 * the device can only be part of the array if
712 * we are reverting a reshape, so this section will
715 if (conf
->raid_disks
>= conf
->previous_raid_disks
)
719 if (conf
->raid_disks
== conf
->previous_raid_disks
)
723 for (i
= 0; i
< conf
->raid_disks
; i
++) {
724 struct md_rdev
*rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
725 if (rdev
&& test_bit(Faulty
, &rdev
->flags
))
726 rdev
= rcu_dereference(conf
->disks
[i
].replacement
);
727 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
))
729 else if (test_bit(In_sync
, &rdev
->flags
))
732 /* not in-sync or faulty.
733 * If reshape increases the number of devices, this
734 * section has already been recovered, else it
735 * almost certainly hasn't.
737 if (conf
->raid_disks
<= conf
->previous_raid_disks
)
741 if (degraded2
> degraded
)
746 static bool has_failed(struct r5conf
*conf
)
748 int degraded
= conf
->mddev
->degraded
;
750 if (test_bit(MD_BROKEN
, &conf
->mddev
->flags
))
753 if (conf
->mddev
->reshape_position
!= MaxSector
)
754 degraded
= raid5_calc_degraded(conf
);
756 return degraded
> conf
->max_degraded
;
762 STRIPE_SCHEDULE_AND_RETRY
,
766 struct stripe_request_ctx
{
767 /* a reference to the last stripe_head for batching */
768 struct stripe_head
*batch_last
;
770 /* first sector in the request */
771 sector_t first_sector
;
773 /* last sector in the request */
774 sector_t last_sector
;
777 * bitmap to track stripe sectors that have been added to stripes
778 * add one to account for unaligned requests
780 DECLARE_BITMAP(sectors_to_do
, RAID5_MAX_REQ_STRIPES
+ 1);
782 /* the request had REQ_PREFLUSH, cleared after the first stripe_head */
787 * Block until another thread clears R5_INACTIVE_BLOCKED or
788 * there are fewer than 3/4 the maximum number of active stripes
789 * and there is an inactive stripe available.
791 static bool is_inactive_blocked(struct r5conf
*conf
, int hash
)
793 if (list_empty(conf
->inactive_list
+ hash
))
796 if (!test_bit(R5_INACTIVE_BLOCKED
, &conf
->cache_state
))
799 return (atomic_read(&conf
->active_stripes
) <
800 (conf
->max_nr_stripes
* 3 / 4));
803 struct stripe_head
*raid5_get_active_stripe(struct r5conf
*conf
,
804 struct stripe_request_ctx
*ctx
, sector_t sector
,
807 struct stripe_head
*sh
;
808 int hash
= stripe_hash_locks_hash(conf
, sector
);
809 int previous
= !!(flags
& R5_GAS_PREVIOUS
);
811 pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector
);
813 spin_lock_irq(conf
->hash_locks
+ hash
);
816 if (!(flags
& R5_GAS_NOQUIESCE
) && conf
->quiesce
) {
818 * Must release the reference to batch_last before
819 * waiting, on quiesce, otherwise the batch_last will
820 * hold a reference to a stripe and raid5_quiesce()
821 * will deadlock waiting for active_stripes to go to
824 if (ctx
&& ctx
->batch_last
) {
825 raid5_release_stripe(ctx
->batch_last
);
826 ctx
->batch_last
= NULL
;
829 wait_event_lock_irq(conf
->wait_for_quiescent
,
831 *(conf
->hash_locks
+ hash
));
834 sh
= find_get_stripe(conf
, sector
, conf
->generation
- previous
,
839 if (!test_bit(R5_INACTIVE_BLOCKED
, &conf
->cache_state
)) {
840 sh
= get_free_stripe(conf
, hash
);
842 r5c_check_stripe_cache_usage(conf
);
843 init_stripe(sh
, sector
, previous
);
844 atomic_inc(&sh
->count
);
848 if (!test_bit(R5_DID_ALLOC
, &conf
->cache_state
))
849 set_bit(R5_ALLOC_MORE
, &conf
->cache_state
);
852 if (flags
& R5_GAS_NOBLOCK
)
855 set_bit(R5_INACTIVE_BLOCKED
, &conf
->cache_state
);
856 r5l_wake_reclaim(conf
->log
, 0);
858 /* release batch_last before wait to avoid risk of deadlock */
859 if (ctx
&& ctx
->batch_last
) {
860 raid5_release_stripe(ctx
->batch_last
);
861 ctx
->batch_last
= NULL
;
864 wait_event_lock_irq(conf
->wait_for_stripe
,
865 is_inactive_blocked(conf
, hash
),
866 *(conf
->hash_locks
+ hash
));
867 clear_bit(R5_INACTIVE_BLOCKED
, &conf
->cache_state
);
870 spin_unlock_irq(conf
->hash_locks
+ hash
);
874 static bool is_full_stripe_write(struct stripe_head
*sh
)
876 BUG_ON(sh
->overwrite_disks
> (sh
->disks
- sh
->raid_conf
->max_degraded
));
877 return sh
->overwrite_disks
== (sh
->disks
- sh
->raid_conf
->max_degraded
);
880 static void lock_two_stripes(struct stripe_head
*sh1
, struct stripe_head
*sh2
)
881 __acquires(&sh1
->stripe_lock
)
882 __acquires(&sh2
->stripe_lock
)
885 spin_lock_irq(&sh2
->stripe_lock
);
886 spin_lock_nested(&sh1
->stripe_lock
, 1);
888 spin_lock_irq(&sh1
->stripe_lock
);
889 spin_lock_nested(&sh2
->stripe_lock
, 1);
893 static void unlock_two_stripes(struct stripe_head
*sh1
, struct stripe_head
*sh2
)
894 __releases(&sh1
->stripe_lock
)
895 __releases(&sh2
->stripe_lock
)
897 spin_unlock(&sh1
->stripe_lock
);
898 spin_unlock_irq(&sh2
->stripe_lock
);
901 /* Only freshly new full stripe normal write stripe can be added to a batch list */
902 static bool stripe_can_batch(struct stripe_head
*sh
)
904 struct r5conf
*conf
= sh
->raid_conf
;
906 if (raid5_has_log(conf
) || raid5_has_ppl(conf
))
908 return test_bit(STRIPE_BATCH_READY
, &sh
->state
) &&
909 !test_bit(STRIPE_BITMAP_PENDING
, &sh
->state
) &&
910 is_full_stripe_write(sh
);
913 /* we only do back search */
914 static void stripe_add_to_batch_list(struct r5conf
*conf
,
915 struct stripe_head
*sh
, struct stripe_head
*last_sh
)
917 struct stripe_head
*head
;
918 sector_t head_sector
, tmp_sec
;
922 /* Don't cross chunks, so stripe pd_idx/qd_idx is the same */
923 tmp_sec
= sh
->sector
;
924 if (!sector_div(tmp_sec
, conf
->chunk_sectors
))
926 head_sector
= sh
->sector
- RAID5_STRIPE_SECTORS(conf
);
928 if (last_sh
&& head_sector
== last_sh
->sector
) {
930 atomic_inc(&head
->count
);
932 hash
= stripe_hash_locks_hash(conf
, head_sector
);
933 spin_lock_irq(conf
->hash_locks
+ hash
);
934 head
= find_get_stripe(conf
, head_sector
, conf
->generation
,
936 spin_unlock_irq(conf
->hash_locks
+ hash
);
939 if (!stripe_can_batch(head
))
943 lock_two_stripes(head
, sh
);
944 /* clear_batch_ready clear the flag */
945 if (!stripe_can_batch(head
) || !stripe_can_batch(sh
))
952 while (dd_idx
== sh
->pd_idx
|| dd_idx
== sh
->qd_idx
)
954 if (head
->dev
[dd_idx
].towrite
->bi_opf
!= sh
->dev
[dd_idx
].towrite
->bi_opf
||
955 bio_op(head
->dev
[dd_idx
].towrite
) != bio_op(sh
->dev
[dd_idx
].towrite
))
958 if (head
->batch_head
) {
959 spin_lock(&head
->batch_head
->batch_lock
);
960 /* This batch list is already running */
961 if (!stripe_can_batch(head
)) {
962 spin_unlock(&head
->batch_head
->batch_lock
);
966 * We must assign batch_head of this stripe within the
967 * batch_lock, otherwise clear_batch_ready of batch head
968 * stripe could clear BATCH_READY bit of this stripe and
969 * this stripe->batch_head doesn't get assigned, which
970 * could confuse clear_batch_ready for this stripe
972 sh
->batch_head
= head
->batch_head
;
975 * at this point, head's BATCH_READY could be cleared, but we
976 * can still add the stripe to batch list
978 list_add(&sh
->batch_list
, &head
->batch_list
);
979 spin_unlock(&head
->batch_head
->batch_lock
);
981 head
->batch_head
= head
;
982 sh
->batch_head
= head
->batch_head
;
983 spin_lock(&head
->batch_lock
);
984 list_add_tail(&sh
->batch_list
, &head
->batch_list
);
985 spin_unlock(&head
->batch_lock
);
988 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
989 if (atomic_dec_return(&conf
->preread_active_stripes
)
991 md_wakeup_thread(conf
->mddev
->thread
);
993 if (test_and_clear_bit(STRIPE_BIT_DELAY
, &sh
->state
)) {
994 int seq
= sh
->bm_seq
;
995 if (test_bit(STRIPE_BIT_DELAY
, &sh
->batch_head
->state
) &&
996 sh
->batch_head
->bm_seq
> seq
)
997 seq
= sh
->batch_head
->bm_seq
;
998 set_bit(STRIPE_BIT_DELAY
, &sh
->batch_head
->state
);
999 sh
->batch_head
->bm_seq
= seq
;
1002 atomic_inc(&sh
->count
);
1004 unlock_two_stripes(head
, sh
);
1006 raid5_release_stripe(head
);
1009 /* Determine if 'data_offset' or 'new_data_offset' should be used
1010 * in this stripe_head.
1012 static int use_new_offset(struct r5conf
*conf
, struct stripe_head
*sh
)
1014 sector_t progress
= conf
->reshape_progress
;
1015 /* Need a memory barrier to make sure we see the value
1016 * of conf->generation, or ->data_offset that was set before
1017 * reshape_progress was updated.
1020 if (progress
== MaxSector
)
1022 if (sh
->generation
== conf
->generation
- 1)
1024 /* We are in a reshape, and this is a new-generation stripe,
1025 * so use new_data_offset.
1030 static void dispatch_bio_list(struct bio_list
*tmp
)
1034 while ((bio
= bio_list_pop(tmp
)))
1035 submit_bio_noacct(bio
);
1038 static int cmp_stripe(void *priv
, const struct list_head
*a
,
1039 const struct list_head
*b
)
1041 const struct r5pending_data
*da
= list_entry(a
,
1042 struct r5pending_data
, sibling
);
1043 const struct r5pending_data
*db
= list_entry(b
,
1044 struct r5pending_data
, sibling
);
1045 if (da
->sector
> db
->sector
)
1047 if (da
->sector
< db
->sector
)
1052 static void dispatch_defer_bios(struct r5conf
*conf
, int target
,
1053 struct bio_list
*list
)
1055 struct r5pending_data
*data
;
1056 struct list_head
*first
, *next
= NULL
;
1059 if (conf
->pending_data_cnt
== 0)
1062 list_sort(NULL
, &conf
->pending_list
, cmp_stripe
);
1064 first
= conf
->pending_list
.next
;
1066 /* temporarily move the head */
1067 if (conf
->next_pending_data
)
1068 list_move_tail(&conf
->pending_list
,
1069 &conf
->next_pending_data
->sibling
);
1071 while (!list_empty(&conf
->pending_list
)) {
1072 data
= list_first_entry(&conf
->pending_list
,
1073 struct r5pending_data
, sibling
);
1074 if (&data
->sibling
== first
)
1075 first
= data
->sibling
.next
;
1076 next
= data
->sibling
.next
;
1078 bio_list_merge(list
, &data
->bios
);
1079 list_move(&data
->sibling
, &conf
->free_list
);
1084 conf
->pending_data_cnt
-= cnt
;
1085 BUG_ON(conf
->pending_data_cnt
< 0 || cnt
< target
);
1087 if (next
!= &conf
->pending_list
)
1088 conf
->next_pending_data
= list_entry(next
,
1089 struct r5pending_data
, sibling
);
1091 conf
->next_pending_data
= NULL
;
1092 /* list isn't empty */
1093 if (first
!= &conf
->pending_list
)
1094 list_move_tail(&conf
->pending_list
, first
);
1097 static void flush_deferred_bios(struct r5conf
*conf
)
1099 struct bio_list tmp
= BIO_EMPTY_LIST
;
1101 if (conf
->pending_data_cnt
== 0)
1104 spin_lock(&conf
->pending_bios_lock
);
1105 dispatch_defer_bios(conf
, conf
->pending_data_cnt
, &tmp
);
1106 BUG_ON(conf
->pending_data_cnt
!= 0);
1107 spin_unlock(&conf
->pending_bios_lock
);
1109 dispatch_bio_list(&tmp
);
1112 static void defer_issue_bios(struct r5conf
*conf
, sector_t sector
,
1113 struct bio_list
*bios
)
1115 struct bio_list tmp
= BIO_EMPTY_LIST
;
1116 struct r5pending_data
*ent
;
1118 spin_lock(&conf
->pending_bios_lock
);
1119 ent
= list_first_entry(&conf
->free_list
, struct r5pending_data
,
1121 list_move_tail(&ent
->sibling
, &conf
->pending_list
);
1122 ent
->sector
= sector
;
1123 bio_list_init(&ent
->bios
);
1124 bio_list_merge(&ent
->bios
, bios
);
1125 conf
->pending_data_cnt
++;
1126 if (conf
->pending_data_cnt
>= PENDING_IO_MAX
)
1127 dispatch_defer_bios(conf
, PENDING_IO_ONE_FLUSH
, &tmp
);
1129 spin_unlock(&conf
->pending_bios_lock
);
1131 dispatch_bio_list(&tmp
);
1135 raid5_end_read_request(struct bio
*bi
);
1137 raid5_end_write_request(struct bio
*bi
);
1139 static void ops_run_io(struct stripe_head
*sh
, struct stripe_head_state
*s
)
1141 struct r5conf
*conf
= sh
->raid_conf
;
1142 int i
, disks
= sh
->disks
;
1143 struct stripe_head
*head_sh
= sh
;
1144 struct bio_list pending_bios
= BIO_EMPTY_LIST
;
1150 if (log_stripe(sh
, s
) == 0)
1153 should_defer
= conf
->batch_bio_dispatch
&& conf
->group_cnt
;
1155 for (i
= disks
; i
--; ) {
1157 blk_opf_t op_flags
= 0;
1158 int replace_only
= 0;
1159 struct bio
*bi
, *rbi
;
1160 struct md_rdev
*rdev
, *rrdev
= NULL
;
1163 if (test_and_clear_bit(R5_Wantwrite
, &sh
->dev
[i
].flags
)) {
1165 if (test_and_clear_bit(R5_WantFUA
, &sh
->dev
[i
].flags
))
1167 if (test_bit(R5_Discard
, &sh
->dev
[i
].flags
))
1168 op
= REQ_OP_DISCARD
;
1169 } else if (test_and_clear_bit(R5_Wantread
, &sh
->dev
[i
].flags
))
1171 else if (test_and_clear_bit(R5_WantReplace
,
1172 &sh
->dev
[i
].flags
)) {
1177 if (test_and_clear_bit(R5_SyncIO
, &sh
->dev
[i
].flags
))
1178 op_flags
|= REQ_SYNC
;
1183 rbi
= &dev
->rreq
; /* For writing to replacement */
1186 rrdev
= rcu_dereference(conf
->disks
[i
].replacement
);
1187 smp_mb(); /* Ensure that if rrdev is NULL, rdev won't be */
1188 rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
1193 if (op_is_write(op
)) {
1197 /* We raced and saw duplicates */
1200 if (test_bit(R5_ReadRepl
, &head_sh
->dev
[i
].flags
) && rrdev
)
1205 if (rdev
&& test_bit(Faulty
, &rdev
->flags
))
1208 atomic_inc(&rdev
->nr_pending
);
1209 if (rrdev
&& test_bit(Faulty
, &rrdev
->flags
))
1212 atomic_inc(&rrdev
->nr_pending
);
1215 /* We have already checked bad blocks for reads. Now
1216 * need to check for writes. We never accept write errors
1217 * on the replacement, so we don't to check rrdev.
1219 while (op_is_write(op
) && rdev
&&
1220 test_bit(WriteErrorSeen
, &rdev
->flags
)) {
1223 int bad
= is_badblock(rdev
, sh
->sector
, RAID5_STRIPE_SECTORS(conf
),
1224 &first_bad
, &bad_sectors
);
1229 set_bit(BlockedBadBlocks
, &rdev
->flags
);
1230 if (!conf
->mddev
->external
&&
1231 conf
->mddev
->sb_flags
) {
1232 /* It is very unlikely, but we might
1233 * still need to write out the
1234 * bad block log - better give it
1236 md_check_recovery(conf
->mddev
);
1239 * Because md_wait_for_blocked_rdev
1240 * will dec nr_pending, we must
1241 * increment it first.
1243 atomic_inc(&rdev
->nr_pending
);
1244 md_wait_for_blocked_rdev(rdev
, conf
->mddev
);
1246 /* Acknowledged bad block - skip the write */
1247 rdev_dec_pending(rdev
, conf
->mddev
);
1253 if (s
->syncing
|| s
->expanding
|| s
->expanded
1255 md_sync_acct(rdev
->bdev
, RAID5_STRIPE_SECTORS(conf
));
1257 set_bit(STRIPE_IO_STARTED
, &sh
->state
);
1259 bio_init(bi
, rdev
->bdev
, &dev
->vec
, 1, op
| op_flags
);
1260 bi
->bi_end_io
= op_is_write(op
)
1261 ? raid5_end_write_request
1262 : raid5_end_read_request
;
1263 bi
->bi_private
= sh
;
1265 pr_debug("%s: for %llu schedule op %d on disc %d\n",
1266 __func__
, (unsigned long long)sh
->sector
,
1268 atomic_inc(&sh
->count
);
1270 atomic_inc(&head_sh
->count
);
1271 if (use_new_offset(conf
, sh
))
1272 bi
->bi_iter
.bi_sector
= (sh
->sector
1273 + rdev
->new_data_offset
);
1275 bi
->bi_iter
.bi_sector
= (sh
->sector
1276 + rdev
->data_offset
);
1277 if (test_bit(R5_ReadNoMerge
, &head_sh
->dev
[i
].flags
))
1278 bi
->bi_opf
|= REQ_NOMERGE
;
1280 if (test_bit(R5_SkipCopy
, &sh
->dev
[i
].flags
))
1281 WARN_ON(test_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
));
1283 if (!op_is_write(op
) &&
1284 test_bit(R5_InJournal
, &sh
->dev
[i
].flags
))
1286 * issuing read for a page in journal, this
1287 * must be preparing for prexor in rmw; read
1288 * the data into orig_page
1290 sh
->dev
[i
].vec
.bv_page
= sh
->dev
[i
].orig_page
;
1292 sh
->dev
[i
].vec
.bv_page
= sh
->dev
[i
].page
;
1294 bi
->bi_io_vec
[0].bv_len
= RAID5_STRIPE_SIZE(conf
);
1295 bi
->bi_io_vec
[0].bv_offset
= sh
->dev
[i
].offset
;
1296 bi
->bi_iter
.bi_size
= RAID5_STRIPE_SIZE(conf
);
1298 * If this is discard request, set bi_vcnt 0. We don't
1299 * want to confuse SCSI because SCSI will replace payload
1301 if (op
== REQ_OP_DISCARD
)
1304 set_bit(R5_DOUBLE_LOCKED
, &sh
->dev
[i
].flags
);
1306 if (conf
->mddev
->gendisk
)
1307 trace_block_bio_remap(bi
,
1308 disk_devt(conf
->mddev
->gendisk
),
1310 if (should_defer
&& op_is_write(op
))
1311 bio_list_add(&pending_bios
, bi
);
1313 submit_bio_noacct(bi
);
1316 if (s
->syncing
|| s
->expanding
|| s
->expanded
1318 md_sync_acct(rrdev
->bdev
, RAID5_STRIPE_SECTORS(conf
));
1320 set_bit(STRIPE_IO_STARTED
, &sh
->state
);
1322 bio_init(rbi
, rrdev
->bdev
, &dev
->rvec
, 1, op
| op_flags
);
1323 BUG_ON(!op_is_write(op
));
1324 rbi
->bi_end_io
= raid5_end_write_request
;
1325 rbi
->bi_private
= sh
;
1327 pr_debug("%s: for %llu schedule op %d on "
1328 "replacement disc %d\n",
1329 __func__
, (unsigned long long)sh
->sector
,
1331 atomic_inc(&sh
->count
);
1333 atomic_inc(&head_sh
->count
);
1334 if (use_new_offset(conf
, sh
))
1335 rbi
->bi_iter
.bi_sector
= (sh
->sector
1336 + rrdev
->new_data_offset
);
1338 rbi
->bi_iter
.bi_sector
= (sh
->sector
1339 + rrdev
->data_offset
);
1340 if (test_bit(R5_SkipCopy
, &sh
->dev
[i
].flags
))
1341 WARN_ON(test_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
));
1342 sh
->dev
[i
].rvec
.bv_page
= sh
->dev
[i
].page
;
1344 rbi
->bi_io_vec
[0].bv_len
= RAID5_STRIPE_SIZE(conf
);
1345 rbi
->bi_io_vec
[0].bv_offset
= sh
->dev
[i
].offset
;
1346 rbi
->bi_iter
.bi_size
= RAID5_STRIPE_SIZE(conf
);
1348 * If this is discard request, set bi_vcnt 0. We don't
1349 * want to confuse SCSI because SCSI will replace payload
1351 if (op
== REQ_OP_DISCARD
)
1353 if (conf
->mddev
->gendisk
)
1354 trace_block_bio_remap(rbi
,
1355 disk_devt(conf
->mddev
->gendisk
),
1357 if (should_defer
&& op_is_write(op
))
1358 bio_list_add(&pending_bios
, rbi
);
1360 submit_bio_noacct(rbi
);
1362 if (!rdev
&& !rrdev
) {
1363 if (op_is_write(op
))
1364 set_bit(STRIPE_DEGRADED
, &sh
->state
);
1365 pr_debug("skip op %d on disc %d for sector %llu\n",
1366 bi
->bi_opf
, i
, (unsigned long long)sh
->sector
);
1367 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
1368 set_bit(STRIPE_HANDLE
, &sh
->state
);
1371 if (!head_sh
->batch_head
)
1373 sh
= list_first_entry(&sh
->batch_list
, struct stripe_head
,
1379 if (should_defer
&& !bio_list_empty(&pending_bios
))
1380 defer_issue_bios(conf
, head_sh
->sector
, &pending_bios
);
1383 static struct dma_async_tx_descriptor
*
1384 async_copy_data(int frombio
, struct bio
*bio
, struct page
**page
,
1385 unsigned int poff
, sector_t sector
, struct dma_async_tx_descriptor
*tx
,
1386 struct stripe_head
*sh
, int no_skipcopy
)
1389 struct bvec_iter iter
;
1390 struct page
*bio_page
;
1392 struct async_submit_ctl submit
;
1393 enum async_tx_flags flags
= 0;
1394 struct r5conf
*conf
= sh
->raid_conf
;
1396 if (bio
->bi_iter
.bi_sector
>= sector
)
1397 page_offset
= (signed)(bio
->bi_iter
.bi_sector
- sector
) * 512;
1399 page_offset
= (signed)(sector
- bio
->bi_iter
.bi_sector
) * -512;
1402 flags
|= ASYNC_TX_FENCE
;
1403 init_async_submit(&submit
, flags
, tx
, NULL
, NULL
, NULL
);
1405 bio_for_each_segment(bvl
, bio
, iter
) {
1406 int len
= bvl
.bv_len
;
1410 if (page_offset
< 0) {
1411 b_offset
= -page_offset
;
1412 page_offset
+= b_offset
;
1416 if (len
> 0 && page_offset
+ len
> RAID5_STRIPE_SIZE(conf
))
1417 clen
= RAID5_STRIPE_SIZE(conf
) - page_offset
;
1422 b_offset
+= bvl
.bv_offset
;
1423 bio_page
= bvl
.bv_page
;
1425 if (conf
->skip_copy
&&
1426 b_offset
== 0 && page_offset
== 0 &&
1427 clen
== RAID5_STRIPE_SIZE(conf
) &&
1431 tx
= async_memcpy(*page
, bio_page
, page_offset
+ poff
,
1432 b_offset
, clen
, &submit
);
1434 tx
= async_memcpy(bio_page
, *page
, b_offset
,
1435 page_offset
+ poff
, clen
, &submit
);
1437 /* chain the operations */
1438 submit
.depend_tx
= tx
;
1440 if (clen
< len
) /* hit end of page */
1448 static void ops_complete_biofill(void *stripe_head_ref
)
1450 struct stripe_head
*sh
= stripe_head_ref
;
1452 struct r5conf
*conf
= sh
->raid_conf
;
1454 pr_debug("%s: stripe %llu\n", __func__
,
1455 (unsigned long long)sh
->sector
);
1457 /* clear completed biofills */
1458 for (i
= sh
->disks
; i
--; ) {
1459 struct r5dev
*dev
= &sh
->dev
[i
];
1461 /* acknowledge completion of a biofill operation */
1462 /* and check if we need to reply to a read request,
1463 * new R5_Wantfill requests are held off until
1464 * !STRIPE_BIOFILL_RUN
1466 if (test_and_clear_bit(R5_Wantfill
, &dev
->flags
)) {
1467 struct bio
*rbi
, *rbi2
;
1472 while (rbi
&& rbi
->bi_iter
.bi_sector
<
1473 dev
->sector
+ RAID5_STRIPE_SECTORS(conf
)) {
1474 rbi2
= r5_next_bio(conf
, rbi
, dev
->sector
);
1480 clear_bit(STRIPE_BIOFILL_RUN
, &sh
->state
);
1482 set_bit(STRIPE_HANDLE
, &sh
->state
);
1483 raid5_release_stripe(sh
);
1486 static void ops_run_biofill(struct stripe_head
*sh
)
1488 struct dma_async_tx_descriptor
*tx
= NULL
;
1489 struct async_submit_ctl submit
;
1491 struct r5conf
*conf
= sh
->raid_conf
;
1493 BUG_ON(sh
->batch_head
);
1494 pr_debug("%s: stripe %llu\n", __func__
,
1495 (unsigned long long)sh
->sector
);
1497 for (i
= sh
->disks
; i
--; ) {
1498 struct r5dev
*dev
= &sh
->dev
[i
];
1499 if (test_bit(R5_Wantfill
, &dev
->flags
)) {
1501 spin_lock_irq(&sh
->stripe_lock
);
1502 dev
->read
= rbi
= dev
->toread
;
1504 spin_unlock_irq(&sh
->stripe_lock
);
1505 while (rbi
&& rbi
->bi_iter
.bi_sector
<
1506 dev
->sector
+ RAID5_STRIPE_SECTORS(conf
)) {
1507 tx
= async_copy_data(0, rbi
, &dev
->page
,
1509 dev
->sector
, tx
, sh
, 0);
1510 rbi
= r5_next_bio(conf
, rbi
, dev
->sector
);
1515 atomic_inc(&sh
->count
);
1516 init_async_submit(&submit
, ASYNC_TX_ACK
, tx
, ops_complete_biofill
, sh
, NULL
);
1517 async_trigger_callback(&submit
);
1520 static void mark_target_uptodate(struct stripe_head
*sh
, int target
)
1527 tgt
= &sh
->dev
[target
];
1528 set_bit(R5_UPTODATE
, &tgt
->flags
);
1529 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
1530 clear_bit(R5_Wantcompute
, &tgt
->flags
);
1533 static void ops_complete_compute(void *stripe_head_ref
)
1535 struct stripe_head
*sh
= stripe_head_ref
;
1537 pr_debug("%s: stripe %llu\n", __func__
,
1538 (unsigned long long)sh
->sector
);
1540 /* mark the computed target(s) as uptodate */
1541 mark_target_uptodate(sh
, sh
->ops
.target
);
1542 mark_target_uptodate(sh
, sh
->ops
.target2
);
1544 clear_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
1545 if (sh
->check_state
== check_state_compute_run
)
1546 sh
->check_state
= check_state_compute_result
;
1547 set_bit(STRIPE_HANDLE
, &sh
->state
);
1548 raid5_release_stripe(sh
);
1551 /* return a pointer to the address conversion region of the scribble buffer */
1552 static struct page
**to_addr_page(struct raid5_percpu
*percpu
, int i
)
1554 return percpu
->scribble
+ i
* percpu
->scribble_obj_size
;
1557 /* return a pointer to the address conversion region of the scribble buffer */
1558 static addr_conv_t
*to_addr_conv(struct stripe_head
*sh
,
1559 struct raid5_percpu
*percpu
, int i
)
1561 return (void *) (to_addr_page(percpu
, i
) + sh
->disks
+ 2);
1565 * Return a pointer to record offset address.
1567 static unsigned int *
1568 to_addr_offs(struct stripe_head
*sh
, struct raid5_percpu
*percpu
)
1570 return (unsigned int *) (to_addr_conv(sh
, percpu
, 0) + sh
->disks
+ 2);
1573 static struct dma_async_tx_descriptor
*
1574 ops_run_compute5(struct stripe_head
*sh
, struct raid5_percpu
*percpu
)
1576 int disks
= sh
->disks
;
1577 struct page
**xor_srcs
= to_addr_page(percpu
, 0);
1578 unsigned int *off_srcs
= to_addr_offs(sh
, percpu
);
1579 int target
= sh
->ops
.target
;
1580 struct r5dev
*tgt
= &sh
->dev
[target
];
1581 struct page
*xor_dest
= tgt
->page
;
1582 unsigned int off_dest
= tgt
->offset
;
1584 struct dma_async_tx_descriptor
*tx
;
1585 struct async_submit_ctl submit
;
1588 BUG_ON(sh
->batch_head
);
1590 pr_debug("%s: stripe %llu block: %d\n",
1591 __func__
, (unsigned long long)sh
->sector
, target
);
1592 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
1594 for (i
= disks
; i
--; ) {
1596 off_srcs
[count
] = sh
->dev
[i
].offset
;
1597 xor_srcs
[count
++] = sh
->dev
[i
].page
;
1601 atomic_inc(&sh
->count
);
1603 init_async_submit(&submit
, ASYNC_TX_FENCE
|ASYNC_TX_XOR_ZERO_DST
, NULL
,
1604 ops_complete_compute
, sh
, to_addr_conv(sh
, percpu
, 0));
1605 if (unlikely(count
== 1))
1606 tx
= async_memcpy(xor_dest
, xor_srcs
[0], off_dest
, off_srcs
[0],
1607 RAID5_STRIPE_SIZE(sh
->raid_conf
), &submit
);
1609 tx
= async_xor_offs(xor_dest
, off_dest
, xor_srcs
, off_srcs
, count
,
1610 RAID5_STRIPE_SIZE(sh
->raid_conf
), &submit
);
1615 /* set_syndrome_sources - populate source buffers for gen_syndrome
1616 * @srcs - (struct page *) array of size sh->disks
1617 * @offs - (unsigned int) array of offset for each page
1618 * @sh - stripe_head to parse
1620 * Populates srcs in proper layout order for the stripe and returns the
1621 * 'count' of sources to be used in a call to async_gen_syndrome. The P
1622 * destination buffer is recorded in srcs[count] and the Q destination
1623 * is recorded in srcs[count+1]].
1625 static int set_syndrome_sources(struct page
**srcs
,
1627 struct stripe_head
*sh
,
1630 int disks
= sh
->disks
;
1631 int syndrome_disks
= sh
->ddf_layout
? disks
: (disks
- 2);
1632 int d0_idx
= raid6_d0(sh
);
1636 for (i
= 0; i
< disks
; i
++)
1642 int slot
= raid6_idx_to_slot(i
, sh
, &count
, syndrome_disks
);
1643 struct r5dev
*dev
= &sh
->dev
[i
];
1645 if (i
== sh
->qd_idx
|| i
== sh
->pd_idx
||
1646 (srctype
== SYNDROME_SRC_ALL
) ||
1647 (srctype
== SYNDROME_SRC_WANT_DRAIN
&&
1648 (test_bit(R5_Wantdrain
, &dev
->flags
) ||
1649 test_bit(R5_InJournal
, &dev
->flags
))) ||
1650 (srctype
== SYNDROME_SRC_WRITTEN
&&
1652 test_bit(R5_InJournal
, &dev
->flags
)))) {
1653 if (test_bit(R5_InJournal
, &dev
->flags
))
1654 srcs
[slot
] = sh
->dev
[i
].orig_page
;
1656 srcs
[slot
] = sh
->dev
[i
].page
;
1658 * For R5_InJournal, PAGE_SIZE must be 4KB and will
1659 * not shared page. In that case, dev[i].offset
1662 offs
[slot
] = sh
->dev
[i
].offset
;
1664 i
= raid6_next_disk(i
, disks
);
1665 } while (i
!= d0_idx
);
1667 return syndrome_disks
;
1670 static struct dma_async_tx_descriptor
*
1671 ops_run_compute6_1(struct stripe_head
*sh
, struct raid5_percpu
*percpu
)
1673 int disks
= sh
->disks
;
1674 struct page
**blocks
= to_addr_page(percpu
, 0);
1675 unsigned int *offs
= to_addr_offs(sh
, percpu
);
1677 int qd_idx
= sh
->qd_idx
;
1678 struct dma_async_tx_descriptor
*tx
;
1679 struct async_submit_ctl submit
;
1682 unsigned int dest_off
;
1686 BUG_ON(sh
->batch_head
);
1687 if (sh
->ops
.target
< 0)
1688 target
= sh
->ops
.target2
;
1689 else if (sh
->ops
.target2
< 0)
1690 target
= sh
->ops
.target
;
1692 /* we should only have one valid target */
1695 pr_debug("%s: stripe %llu block: %d\n",
1696 __func__
, (unsigned long long)sh
->sector
, target
);
1698 tgt
= &sh
->dev
[target
];
1699 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
1701 dest_off
= tgt
->offset
;
1703 atomic_inc(&sh
->count
);
1705 if (target
== qd_idx
) {
1706 count
= set_syndrome_sources(blocks
, offs
, sh
, SYNDROME_SRC_ALL
);
1707 blocks
[count
] = NULL
; /* regenerating p is not necessary */
1708 BUG_ON(blocks
[count
+1] != dest
); /* q should already be set */
1709 init_async_submit(&submit
, ASYNC_TX_FENCE
, NULL
,
1710 ops_complete_compute
, sh
,
1711 to_addr_conv(sh
, percpu
, 0));
1712 tx
= async_gen_syndrome(blocks
, offs
, count
+2,
1713 RAID5_STRIPE_SIZE(sh
->raid_conf
), &submit
);
1715 /* Compute any data- or p-drive using XOR */
1717 for (i
= disks
; i
-- ; ) {
1718 if (i
== target
|| i
== qd_idx
)
1720 offs
[count
] = sh
->dev
[i
].offset
;
1721 blocks
[count
++] = sh
->dev
[i
].page
;
1724 init_async_submit(&submit
, ASYNC_TX_FENCE
|ASYNC_TX_XOR_ZERO_DST
,
1725 NULL
, ops_complete_compute
, sh
,
1726 to_addr_conv(sh
, percpu
, 0));
1727 tx
= async_xor_offs(dest
, dest_off
, blocks
, offs
, count
,
1728 RAID5_STRIPE_SIZE(sh
->raid_conf
), &submit
);
1734 static struct dma_async_tx_descriptor
*
1735 ops_run_compute6_2(struct stripe_head
*sh
, struct raid5_percpu
*percpu
)
1737 int i
, count
, disks
= sh
->disks
;
1738 int syndrome_disks
= sh
->ddf_layout
? disks
: disks
-2;
1739 int d0_idx
= raid6_d0(sh
);
1740 int faila
= -1, failb
= -1;
1741 int target
= sh
->ops
.target
;
1742 int target2
= sh
->ops
.target2
;
1743 struct r5dev
*tgt
= &sh
->dev
[target
];
1744 struct r5dev
*tgt2
= &sh
->dev
[target2
];
1745 struct dma_async_tx_descriptor
*tx
;
1746 struct page
**blocks
= to_addr_page(percpu
, 0);
1747 unsigned int *offs
= to_addr_offs(sh
, percpu
);
1748 struct async_submit_ctl submit
;
1750 BUG_ON(sh
->batch_head
);
1751 pr_debug("%s: stripe %llu block1: %d block2: %d\n",
1752 __func__
, (unsigned long long)sh
->sector
, target
, target2
);
1753 BUG_ON(target
< 0 || target2
< 0);
1754 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
1755 BUG_ON(!test_bit(R5_Wantcompute
, &tgt2
->flags
));
1757 /* we need to open-code set_syndrome_sources to handle the
1758 * slot number conversion for 'faila' and 'failb'
1760 for (i
= 0; i
< disks
; i
++) {
1767 int slot
= raid6_idx_to_slot(i
, sh
, &count
, syndrome_disks
);
1769 offs
[slot
] = sh
->dev
[i
].offset
;
1770 blocks
[slot
] = sh
->dev
[i
].page
;
1776 i
= raid6_next_disk(i
, disks
);
1777 } while (i
!= d0_idx
);
1779 BUG_ON(faila
== failb
);
1782 pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
1783 __func__
, (unsigned long long)sh
->sector
, faila
, failb
);
1785 atomic_inc(&sh
->count
);
1787 if (failb
== syndrome_disks
+1) {
1788 /* Q disk is one of the missing disks */
1789 if (faila
== syndrome_disks
) {
1790 /* Missing P+Q, just recompute */
1791 init_async_submit(&submit
, ASYNC_TX_FENCE
, NULL
,
1792 ops_complete_compute
, sh
,
1793 to_addr_conv(sh
, percpu
, 0));
1794 return async_gen_syndrome(blocks
, offs
, syndrome_disks
+2,
1795 RAID5_STRIPE_SIZE(sh
->raid_conf
),
1799 unsigned int dest_off
;
1801 int qd_idx
= sh
->qd_idx
;
1803 /* Missing D+Q: recompute D from P, then recompute Q */
1804 if (target
== qd_idx
)
1805 data_target
= target2
;
1807 data_target
= target
;
1810 for (i
= disks
; i
-- ; ) {
1811 if (i
== data_target
|| i
== qd_idx
)
1813 offs
[count
] = sh
->dev
[i
].offset
;
1814 blocks
[count
++] = sh
->dev
[i
].page
;
1816 dest
= sh
->dev
[data_target
].page
;
1817 dest_off
= sh
->dev
[data_target
].offset
;
1818 init_async_submit(&submit
,
1819 ASYNC_TX_FENCE
|ASYNC_TX_XOR_ZERO_DST
,
1821 to_addr_conv(sh
, percpu
, 0));
1822 tx
= async_xor_offs(dest
, dest_off
, blocks
, offs
, count
,
1823 RAID5_STRIPE_SIZE(sh
->raid_conf
),
1826 count
= set_syndrome_sources(blocks
, offs
, sh
, SYNDROME_SRC_ALL
);
1827 init_async_submit(&submit
, ASYNC_TX_FENCE
, tx
,
1828 ops_complete_compute
, sh
,
1829 to_addr_conv(sh
, percpu
, 0));
1830 return async_gen_syndrome(blocks
, offs
, count
+2,
1831 RAID5_STRIPE_SIZE(sh
->raid_conf
),
1835 init_async_submit(&submit
, ASYNC_TX_FENCE
, NULL
,
1836 ops_complete_compute
, sh
,
1837 to_addr_conv(sh
, percpu
, 0));
1838 if (failb
== syndrome_disks
) {
1839 /* We're missing D+P. */
1840 return async_raid6_datap_recov(syndrome_disks
+2,
1841 RAID5_STRIPE_SIZE(sh
->raid_conf
),
1843 blocks
, offs
, &submit
);
1845 /* We're missing D+D. */
1846 return async_raid6_2data_recov(syndrome_disks
+2,
1847 RAID5_STRIPE_SIZE(sh
->raid_conf
),
1849 blocks
, offs
, &submit
);
1854 static void ops_complete_prexor(void *stripe_head_ref
)
1856 struct stripe_head
*sh
= stripe_head_ref
;
1858 pr_debug("%s: stripe %llu\n", __func__
,
1859 (unsigned long long)sh
->sector
);
1861 if (r5c_is_writeback(sh
->raid_conf
->log
))
1863 * raid5-cache write back uses orig_page during prexor.
1864 * After prexor, it is time to free orig_page
1866 r5c_release_extra_page(sh
);
1869 static struct dma_async_tx_descriptor
*
1870 ops_run_prexor5(struct stripe_head
*sh
, struct raid5_percpu
*percpu
,
1871 struct dma_async_tx_descriptor
*tx
)
1873 int disks
= sh
->disks
;
1874 struct page
**xor_srcs
= to_addr_page(percpu
, 0);
1875 unsigned int *off_srcs
= to_addr_offs(sh
, percpu
);
1876 int count
= 0, pd_idx
= sh
->pd_idx
, i
;
1877 struct async_submit_ctl submit
;
1879 /* existing parity data subtracted */
1880 unsigned int off_dest
= off_srcs
[count
] = sh
->dev
[pd_idx
].offset
;
1881 struct page
*xor_dest
= xor_srcs
[count
++] = sh
->dev
[pd_idx
].page
;
1883 BUG_ON(sh
->batch_head
);
1884 pr_debug("%s: stripe %llu\n", __func__
,
1885 (unsigned long long)sh
->sector
);
1887 for (i
= disks
; i
--; ) {
1888 struct r5dev
*dev
= &sh
->dev
[i
];
1889 /* Only process blocks that are known to be uptodate */
1890 if (test_bit(R5_InJournal
, &dev
->flags
)) {
1892 * For this case, PAGE_SIZE must be equal to 4KB and
1893 * page offset is zero.
1895 off_srcs
[count
] = dev
->offset
;
1896 xor_srcs
[count
++] = dev
->orig_page
;
1897 } else if (test_bit(R5_Wantdrain
, &dev
->flags
)) {
1898 off_srcs
[count
] = dev
->offset
;
1899 xor_srcs
[count
++] = dev
->page
;
1903 init_async_submit(&submit
, ASYNC_TX_FENCE
|ASYNC_TX_XOR_DROP_DST
, tx
,
1904 ops_complete_prexor
, sh
, to_addr_conv(sh
, percpu
, 0));
1905 tx
= async_xor_offs(xor_dest
, off_dest
, xor_srcs
, off_srcs
, count
,
1906 RAID5_STRIPE_SIZE(sh
->raid_conf
), &submit
);
1911 static struct dma_async_tx_descriptor
*
1912 ops_run_prexor6(struct stripe_head
*sh
, struct raid5_percpu
*percpu
,
1913 struct dma_async_tx_descriptor
*tx
)
1915 struct page
**blocks
= to_addr_page(percpu
, 0);
1916 unsigned int *offs
= to_addr_offs(sh
, percpu
);
1918 struct async_submit_ctl submit
;
1920 pr_debug("%s: stripe %llu\n", __func__
,
1921 (unsigned long long)sh
->sector
);
1923 count
= set_syndrome_sources(blocks
, offs
, sh
, SYNDROME_SRC_WANT_DRAIN
);
1925 init_async_submit(&submit
, ASYNC_TX_FENCE
|ASYNC_TX_PQ_XOR_DST
, tx
,
1926 ops_complete_prexor
, sh
, to_addr_conv(sh
, percpu
, 0));
1927 tx
= async_gen_syndrome(blocks
, offs
, count
+2,
1928 RAID5_STRIPE_SIZE(sh
->raid_conf
), &submit
);
1933 static struct dma_async_tx_descriptor
*
1934 ops_run_biodrain(struct stripe_head
*sh
, struct dma_async_tx_descriptor
*tx
)
1936 struct r5conf
*conf
= sh
->raid_conf
;
1937 int disks
= sh
->disks
;
1939 struct stripe_head
*head_sh
= sh
;
1941 pr_debug("%s: stripe %llu\n", __func__
,
1942 (unsigned long long)sh
->sector
);
1944 for (i
= disks
; i
--; ) {
1949 if (test_and_clear_bit(R5_Wantdrain
, &head_sh
->dev
[i
].flags
)) {
1955 * clear R5_InJournal, so when rewriting a page in
1956 * journal, it is not skipped by r5l_log_stripe()
1958 clear_bit(R5_InJournal
, &dev
->flags
);
1959 spin_lock_irq(&sh
->stripe_lock
);
1960 chosen
= dev
->towrite
;
1961 dev
->towrite
= NULL
;
1962 sh
->overwrite_disks
= 0;
1963 BUG_ON(dev
->written
);
1964 wbi
= dev
->written
= chosen
;
1965 spin_unlock_irq(&sh
->stripe_lock
);
1966 WARN_ON(dev
->page
!= dev
->orig_page
);
1968 while (wbi
&& wbi
->bi_iter
.bi_sector
<
1969 dev
->sector
+ RAID5_STRIPE_SECTORS(conf
)) {
1970 if (wbi
->bi_opf
& REQ_FUA
)
1971 set_bit(R5_WantFUA
, &dev
->flags
);
1972 if (wbi
->bi_opf
& REQ_SYNC
)
1973 set_bit(R5_SyncIO
, &dev
->flags
);
1974 if (bio_op(wbi
) == REQ_OP_DISCARD
)
1975 set_bit(R5_Discard
, &dev
->flags
);
1977 tx
= async_copy_data(1, wbi
, &dev
->page
,
1979 dev
->sector
, tx
, sh
,
1980 r5c_is_writeback(conf
->log
));
1981 if (dev
->page
!= dev
->orig_page
&&
1982 !r5c_is_writeback(conf
->log
)) {
1983 set_bit(R5_SkipCopy
, &dev
->flags
);
1984 clear_bit(R5_UPTODATE
, &dev
->flags
);
1985 clear_bit(R5_OVERWRITE
, &dev
->flags
);
1988 wbi
= r5_next_bio(conf
, wbi
, dev
->sector
);
1991 if (head_sh
->batch_head
) {
1992 sh
= list_first_entry(&sh
->batch_list
,
2005 static void ops_complete_reconstruct(void *stripe_head_ref
)
2007 struct stripe_head
*sh
= stripe_head_ref
;
2008 int disks
= sh
->disks
;
2009 int pd_idx
= sh
->pd_idx
;
2010 int qd_idx
= sh
->qd_idx
;
2012 bool fua
= false, sync
= false, discard
= false;
2014 pr_debug("%s: stripe %llu\n", __func__
,
2015 (unsigned long long)sh
->sector
);
2017 for (i
= disks
; i
--; ) {
2018 fua
|= test_bit(R5_WantFUA
, &sh
->dev
[i
].flags
);
2019 sync
|= test_bit(R5_SyncIO
, &sh
->dev
[i
].flags
);
2020 discard
|= test_bit(R5_Discard
, &sh
->dev
[i
].flags
);
2023 for (i
= disks
; i
--; ) {
2024 struct r5dev
*dev
= &sh
->dev
[i
];
2026 if (dev
->written
|| i
== pd_idx
|| i
== qd_idx
) {
2027 if (!discard
&& !test_bit(R5_SkipCopy
, &dev
->flags
)) {
2028 set_bit(R5_UPTODATE
, &dev
->flags
);
2029 if (test_bit(STRIPE_EXPAND_READY
, &sh
->state
))
2030 set_bit(R5_Expanded
, &dev
->flags
);
2033 set_bit(R5_WantFUA
, &dev
->flags
);
2035 set_bit(R5_SyncIO
, &dev
->flags
);
2039 if (sh
->reconstruct_state
== reconstruct_state_drain_run
)
2040 sh
->reconstruct_state
= reconstruct_state_drain_result
;
2041 else if (sh
->reconstruct_state
== reconstruct_state_prexor_drain_run
)
2042 sh
->reconstruct_state
= reconstruct_state_prexor_drain_result
;
2044 BUG_ON(sh
->reconstruct_state
!= reconstruct_state_run
);
2045 sh
->reconstruct_state
= reconstruct_state_result
;
2048 set_bit(STRIPE_HANDLE
, &sh
->state
);
2049 raid5_release_stripe(sh
);
2053 ops_run_reconstruct5(struct stripe_head
*sh
, struct raid5_percpu
*percpu
,
2054 struct dma_async_tx_descriptor
*tx
)
2056 int disks
= sh
->disks
;
2057 struct page
**xor_srcs
;
2058 unsigned int *off_srcs
;
2059 struct async_submit_ctl submit
;
2060 int count
, pd_idx
= sh
->pd_idx
, i
;
2061 struct page
*xor_dest
;
2062 unsigned int off_dest
;
2064 unsigned long flags
;
2066 struct stripe_head
*head_sh
= sh
;
2069 pr_debug("%s: stripe %llu\n", __func__
,
2070 (unsigned long long)sh
->sector
);
2072 for (i
= 0; i
< sh
->disks
; i
++) {
2075 if (!test_bit(R5_Discard
, &sh
->dev
[i
].flags
))
2078 if (i
>= sh
->disks
) {
2079 atomic_inc(&sh
->count
);
2080 set_bit(R5_Discard
, &sh
->dev
[pd_idx
].flags
);
2081 ops_complete_reconstruct(sh
);
2086 xor_srcs
= to_addr_page(percpu
, j
);
2087 off_srcs
= to_addr_offs(sh
, percpu
);
2088 /* check if prexor is active which means only process blocks
2089 * that are part of a read-modify-write (written)
2091 if (head_sh
->reconstruct_state
== reconstruct_state_prexor_drain_run
) {
2093 off_dest
= off_srcs
[count
] = sh
->dev
[pd_idx
].offset
;
2094 xor_dest
= xor_srcs
[count
++] = sh
->dev
[pd_idx
].page
;
2095 for (i
= disks
; i
--; ) {
2096 struct r5dev
*dev
= &sh
->dev
[i
];
2097 if (head_sh
->dev
[i
].written
||
2098 test_bit(R5_InJournal
, &head_sh
->dev
[i
].flags
)) {
2099 off_srcs
[count
] = dev
->offset
;
2100 xor_srcs
[count
++] = dev
->page
;
2104 xor_dest
= sh
->dev
[pd_idx
].page
;
2105 off_dest
= sh
->dev
[pd_idx
].offset
;
2106 for (i
= disks
; i
--; ) {
2107 struct r5dev
*dev
= &sh
->dev
[i
];
2109 off_srcs
[count
] = dev
->offset
;
2110 xor_srcs
[count
++] = dev
->page
;
2115 /* 1/ if we prexor'd then the dest is reused as a source
2116 * 2/ if we did not prexor then we are redoing the parity
2117 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
2118 * for the synchronous xor case
2120 last_stripe
= !head_sh
->batch_head
||
2121 list_first_entry(&sh
->batch_list
,
2122 struct stripe_head
, batch_list
) == head_sh
;
2124 flags
= ASYNC_TX_ACK
|
2125 (prexor
? ASYNC_TX_XOR_DROP_DST
: ASYNC_TX_XOR_ZERO_DST
);
2127 atomic_inc(&head_sh
->count
);
2128 init_async_submit(&submit
, flags
, tx
, ops_complete_reconstruct
, head_sh
,
2129 to_addr_conv(sh
, percpu
, j
));
2131 flags
= prexor
? ASYNC_TX_XOR_DROP_DST
: ASYNC_TX_XOR_ZERO_DST
;
2132 init_async_submit(&submit
, flags
, tx
, NULL
, NULL
,
2133 to_addr_conv(sh
, percpu
, j
));
2136 if (unlikely(count
== 1))
2137 tx
= async_memcpy(xor_dest
, xor_srcs
[0], off_dest
, off_srcs
[0],
2138 RAID5_STRIPE_SIZE(sh
->raid_conf
), &submit
);
2140 tx
= async_xor_offs(xor_dest
, off_dest
, xor_srcs
, off_srcs
, count
,
2141 RAID5_STRIPE_SIZE(sh
->raid_conf
), &submit
);
2144 sh
= list_first_entry(&sh
->batch_list
, struct stripe_head
,
2151 ops_run_reconstruct6(struct stripe_head
*sh
, struct raid5_percpu
*percpu
,
2152 struct dma_async_tx_descriptor
*tx
)
2154 struct async_submit_ctl submit
;
2155 struct page
**blocks
;
2157 int count
, i
, j
= 0;
2158 struct stripe_head
*head_sh
= sh
;
2161 unsigned long txflags
;
2163 pr_debug("%s: stripe %llu\n", __func__
, (unsigned long long)sh
->sector
);
2165 for (i
= 0; i
< sh
->disks
; i
++) {
2166 if (sh
->pd_idx
== i
|| sh
->qd_idx
== i
)
2168 if (!test_bit(R5_Discard
, &sh
->dev
[i
].flags
))
2171 if (i
>= sh
->disks
) {
2172 atomic_inc(&sh
->count
);
2173 set_bit(R5_Discard
, &sh
->dev
[sh
->pd_idx
].flags
);
2174 set_bit(R5_Discard
, &sh
->dev
[sh
->qd_idx
].flags
);
2175 ops_complete_reconstruct(sh
);
2180 blocks
= to_addr_page(percpu
, j
);
2181 offs
= to_addr_offs(sh
, percpu
);
2183 if (sh
->reconstruct_state
== reconstruct_state_prexor_drain_run
) {
2184 synflags
= SYNDROME_SRC_WRITTEN
;
2185 txflags
= ASYNC_TX_ACK
| ASYNC_TX_PQ_XOR_DST
;
2187 synflags
= SYNDROME_SRC_ALL
;
2188 txflags
= ASYNC_TX_ACK
;
2191 count
= set_syndrome_sources(blocks
, offs
, sh
, synflags
);
2192 last_stripe
= !head_sh
->batch_head
||
2193 list_first_entry(&sh
->batch_list
,
2194 struct stripe_head
, batch_list
) == head_sh
;
2197 atomic_inc(&head_sh
->count
);
2198 init_async_submit(&submit
, txflags
, tx
, ops_complete_reconstruct
,
2199 head_sh
, to_addr_conv(sh
, percpu
, j
));
2201 init_async_submit(&submit
, 0, tx
, NULL
, NULL
,
2202 to_addr_conv(sh
, percpu
, j
));
2203 tx
= async_gen_syndrome(blocks
, offs
, count
+2,
2204 RAID5_STRIPE_SIZE(sh
->raid_conf
), &submit
);
2207 sh
= list_first_entry(&sh
->batch_list
, struct stripe_head
,
2213 static void ops_complete_check(void *stripe_head_ref
)
2215 struct stripe_head
*sh
= stripe_head_ref
;
2217 pr_debug("%s: stripe %llu\n", __func__
,
2218 (unsigned long long)sh
->sector
);
2220 sh
->check_state
= check_state_check_result
;
2221 set_bit(STRIPE_HANDLE
, &sh
->state
);
2222 raid5_release_stripe(sh
);
2225 static void ops_run_check_p(struct stripe_head
*sh
, struct raid5_percpu
*percpu
)
2227 int disks
= sh
->disks
;
2228 int pd_idx
= sh
->pd_idx
;
2229 int qd_idx
= sh
->qd_idx
;
2230 struct page
*xor_dest
;
2231 unsigned int off_dest
;
2232 struct page
**xor_srcs
= to_addr_page(percpu
, 0);
2233 unsigned int *off_srcs
= to_addr_offs(sh
, percpu
);
2234 struct dma_async_tx_descriptor
*tx
;
2235 struct async_submit_ctl submit
;
2239 pr_debug("%s: stripe %llu\n", __func__
,
2240 (unsigned long long)sh
->sector
);
2242 BUG_ON(sh
->batch_head
);
2244 xor_dest
= sh
->dev
[pd_idx
].page
;
2245 off_dest
= sh
->dev
[pd_idx
].offset
;
2246 off_srcs
[count
] = off_dest
;
2247 xor_srcs
[count
++] = xor_dest
;
2248 for (i
= disks
; i
--; ) {
2249 if (i
== pd_idx
|| i
== qd_idx
)
2251 off_srcs
[count
] = sh
->dev
[i
].offset
;
2252 xor_srcs
[count
++] = sh
->dev
[i
].page
;
2255 init_async_submit(&submit
, 0, NULL
, NULL
, NULL
,
2256 to_addr_conv(sh
, percpu
, 0));
2257 tx
= async_xor_val_offs(xor_dest
, off_dest
, xor_srcs
, off_srcs
, count
,
2258 RAID5_STRIPE_SIZE(sh
->raid_conf
),
2259 &sh
->ops
.zero_sum_result
, &submit
);
2261 atomic_inc(&sh
->count
);
2262 init_async_submit(&submit
, ASYNC_TX_ACK
, tx
, ops_complete_check
, sh
, NULL
);
2263 tx
= async_trigger_callback(&submit
);
2266 static void ops_run_check_pq(struct stripe_head
*sh
, struct raid5_percpu
*percpu
, int checkp
)
2268 struct page
**srcs
= to_addr_page(percpu
, 0);
2269 unsigned int *offs
= to_addr_offs(sh
, percpu
);
2270 struct async_submit_ctl submit
;
2273 pr_debug("%s: stripe %llu checkp: %d\n", __func__
,
2274 (unsigned long long)sh
->sector
, checkp
);
2276 BUG_ON(sh
->batch_head
);
2277 count
= set_syndrome_sources(srcs
, offs
, sh
, SYNDROME_SRC_ALL
);
2281 atomic_inc(&sh
->count
);
2282 init_async_submit(&submit
, ASYNC_TX_ACK
, NULL
, ops_complete_check
,
2283 sh
, to_addr_conv(sh
, percpu
, 0));
2284 async_syndrome_val(srcs
, offs
, count
+2,
2285 RAID5_STRIPE_SIZE(sh
->raid_conf
),
2286 &sh
->ops
.zero_sum_result
, percpu
->spare_page
, 0, &submit
);
2289 static void raid_run_ops(struct stripe_head
*sh
, unsigned long ops_request
)
2291 int overlap_clear
= 0, i
, disks
= sh
->disks
;
2292 struct dma_async_tx_descriptor
*tx
= NULL
;
2293 struct r5conf
*conf
= sh
->raid_conf
;
2294 int level
= conf
->level
;
2295 struct raid5_percpu
*percpu
;
2297 local_lock(&conf
->percpu
->lock
);
2298 percpu
= this_cpu_ptr(conf
->percpu
);
2299 if (test_bit(STRIPE_OP_BIOFILL
, &ops_request
)) {
2300 ops_run_biofill(sh
);
2304 if (test_bit(STRIPE_OP_COMPUTE_BLK
, &ops_request
)) {
2306 tx
= ops_run_compute5(sh
, percpu
);
2308 if (sh
->ops
.target2
< 0 || sh
->ops
.target
< 0)
2309 tx
= ops_run_compute6_1(sh
, percpu
);
2311 tx
= ops_run_compute6_2(sh
, percpu
);
2313 /* terminate the chain if reconstruct is not set to be run */
2314 if (tx
&& !test_bit(STRIPE_OP_RECONSTRUCT
, &ops_request
))
2318 if (test_bit(STRIPE_OP_PREXOR
, &ops_request
)) {
2320 tx
= ops_run_prexor5(sh
, percpu
, tx
);
2322 tx
= ops_run_prexor6(sh
, percpu
, tx
);
2325 if (test_bit(STRIPE_OP_PARTIAL_PARITY
, &ops_request
))
2326 tx
= ops_run_partial_parity(sh
, percpu
, tx
);
2328 if (test_bit(STRIPE_OP_BIODRAIN
, &ops_request
)) {
2329 tx
= ops_run_biodrain(sh
, tx
);
2333 if (test_bit(STRIPE_OP_RECONSTRUCT
, &ops_request
)) {
2335 ops_run_reconstruct5(sh
, percpu
, tx
);
2337 ops_run_reconstruct6(sh
, percpu
, tx
);
2340 if (test_bit(STRIPE_OP_CHECK
, &ops_request
)) {
2341 if (sh
->check_state
== check_state_run
)
2342 ops_run_check_p(sh
, percpu
);
2343 else if (sh
->check_state
== check_state_run_q
)
2344 ops_run_check_pq(sh
, percpu
, 0);
2345 else if (sh
->check_state
== check_state_run_pq
)
2346 ops_run_check_pq(sh
, percpu
, 1);
2351 if (overlap_clear
&& !sh
->batch_head
) {
2352 for (i
= disks
; i
--; ) {
2353 struct r5dev
*dev
= &sh
->dev
[i
];
2354 if (test_and_clear_bit(R5_Overlap
, &dev
->flags
))
2355 wake_up(&sh
->raid_conf
->wait_for_overlap
);
2358 local_unlock(&conf
->percpu
->lock
);
2361 static void free_stripe(struct kmem_cache
*sc
, struct stripe_head
*sh
)
2363 #if PAGE_SIZE != DEFAULT_STRIPE_SIZE
2367 __free_page(sh
->ppl_page
);
2368 kmem_cache_free(sc
, sh
);
2371 static struct stripe_head
*alloc_stripe(struct kmem_cache
*sc
, gfp_t gfp
,
2372 int disks
, struct r5conf
*conf
)
2374 struct stripe_head
*sh
;
2376 sh
= kmem_cache_zalloc(sc
, gfp
);
2378 spin_lock_init(&sh
->stripe_lock
);
2379 spin_lock_init(&sh
->batch_lock
);
2380 INIT_LIST_HEAD(&sh
->batch_list
);
2381 INIT_LIST_HEAD(&sh
->lru
);
2382 INIT_LIST_HEAD(&sh
->r5c
);
2383 INIT_LIST_HEAD(&sh
->log_list
);
2384 atomic_set(&sh
->count
, 1);
2385 sh
->raid_conf
= conf
;
2386 sh
->log_start
= MaxSector
;
2388 if (raid5_has_ppl(conf
)) {
2389 sh
->ppl_page
= alloc_page(gfp
);
2390 if (!sh
->ppl_page
) {
2391 free_stripe(sc
, sh
);
2395 #if PAGE_SIZE != DEFAULT_STRIPE_SIZE
2396 if (init_stripe_shared_pages(sh
, conf
, disks
)) {
2397 free_stripe(sc
, sh
);
2404 static int grow_one_stripe(struct r5conf
*conf
, gfp_t gfp
)
2406 struct stripe_head
*sh
;
2408 sh
= alloc_stripe(conf
->slab_cache
, gfp
, conf
->pool_size
, conf
);
2412 if (grow_buffers(sh
, gfp
)) {
2414 free_stripe(conf
->slab_cache
, sh
);
2417 sh
->hash_lock_index
=
2418 conf
->max_nr_stripes
% NR_STRIPE_HASH_LOCKS
;
2419 /* we just created an active stripe so... */
2420 atomic_inc(&conf
->active_stripes
);
2422 raid5_release_stripe(sh
);
2423 conf
->max_nr_stripes
++;
2427 static int grow_stripes(struct r5conf
*conf
, int num
)
2429 struct kmem_cache
*sc
;
2430 size_t namelen
= sizeof(conf
->cache_name
[0]);
2431 int devs
= max(conf
->raid_disks
, conf
->previous_raid_disks
);
2433 if (conf
->mddev
->gendisk
)
2434 snprintf(conf
->cache_name
[0], namelen
,
2435 "raid%d-%s", conf
->level
, mdname(conf
->mddev
));
2437 snprintf(conf
->cache_name
[0], namelen
,
2438 "raid%d-%p", conf
->level
, conf
->mddev
);
2439 snprintf(conf
->cache_name
[1], namelen
, "%.27s-alt", conf
->cache_name
[0]);
2441 conf
->active_name
= 0;
2442 sc
= kmem_cache_create(conf
->cache_name
[conf
->active_name
],
2443 struct_size_t(struct stripe_head
, dev
, devs
),
2447 conf
->slab_cache
= sc
;
2448 conf
->pool_size
= devs
;
2450 if (!grow_one_stripe(conf
, GFP_KERNEL
))
2457 * scribble_alloc - allocate percpu scribble buffer for required size
2458 * of the scribble region
2459 * @percpu: from for_each_present_cpu() of the caller
2460 * @num: total number of disks in the array
2461 * @cnt: scribble objs count for required size of the scribble region
2463 * The scribble buffer size must be enough to contain:
2464 * 1/ a struct page pointer for each device in the array +2
2465 * 2/ room to convert each entry in (1) to its corresponding dma
2466 * (dma_map_page()) or page (page_address()) address.
2468 * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
2469 * calculate over all devices (not just the data blocks), using zeros in place
2470 * of the P and Q blocks.
2472 static int scribble_alloc(struct raid5_percpu
*percpu
,
2476 sizeof(struct page
*) * (num
+ 2) +
2477 sizeof(addr_conv_t
) * (num
+ 2) +
2478 sizeof(unsigned int) * (num
+ 2);
2482 * If here is in raid array suspend context, it is in memalloc noio
2483 * context as well, there is no potential recursive memory reclaim
2484 * I/Os with the GFP_KERNEL flag.
2486 scribble
= kvmalloc_array(cnt
, obj_size
, GFP_KERNEL
);
2490 kvfree(percpu
->scribble
);
2492 percpu
->scribble
= scribble
;
2493 percpu
->scribble_obj_size
= obj_size
;
2497 static int resize_chunks(struct r5conf
*conf
, int new_disks
, int new_sectors
)
2503 * Never shrink. And mddev_suspend() could deadlock if this is called
2504 * from raid5d. In that case, scribble_disks and scribble_sectors
2505 * should equal to new_disks and new_sectors
2507 if (conf
->scribble_disks
>= new_disks
&&
2508 conf
->scribble_sectors
>= new_sectors
)
2510 mddev_suspend(conf
->mddev
);
2513 for_each_present_cpu(cpu
) {
2514 struct raid5_percpu
*percpu
;
2516 percpu
= per_cpu_ptr(conf
->percpu
, cpu
);
2517 err
= scribble_alloc(percpu
, new_disks
,
2518 new_sectors
/ RAID5_STRIPE_SECTORS(conf
));
2524 mddev_resume(conf
->mddev
);
2526 conf
->scribble_disks
= new_disks
;
2527 conf
->scribble_sectors
= new_sectors
;
2532 static int resize_stripes(struct r5conf
*conf
, int newsize
)
2534 /* Make all the stripes able to hold 'newsize' devices.
2535 * New slots in each stripe get 'page' set to a new page.
2537 * This happens in stages:
2538 * 1/ create a new kmem_cache and allocate the required number of
2540 * 2/ gather all the old stripe_heads and transfer the pages across
2541 * to the new stripe_heads. This will have the side effect of
2542 * freezing the array as once all stripe_heads have been collected,
2543 * no IO will be possible. Old stripe heads are freed once their
2544 * pages have been transferred over, and the old kmem_cache is
2545 * freed when all stripes are done.
2546 * 3/ reallocate conf->disks to be suitable bigger. If this fails,
2547 * we simple return a failure status - no need to clean anything up.
2548 * 4/ allocate new pages for the new slots in the new stripe_heads.
2549 * If this fails, we don't bother trying the shrink the
2550 * stripe_heads down again, we just leave them as they are.
2551 * As each stripe_head is processed the new one is released into
2554 * Once step2 is started, we cannot afford to wait for a write,
2555 * so we use GFP_NOIO allocations.
2557 struct stripe_head
*osh
, *nsh
;
2558 LIST_HEAD(newstripes
);
2559 struct disk_info
*ndisks
;
2561 struct kmem_cache
*sc
;
2565 md_allow_write(conf
->mddev
);
2568 sc
= kmem_cache_create(conf
->cache_name
[1-conf
->active_name
],
2569 struct_size_t(struct stripe_head
, dev
, newsize
),
2574 /* Need to ensure auto-resizing doesn't interfere */
2575 mutex_lock(&conf
->cache_size_mutex
);
2577 for (i
= conf
->max_nr_stripes
; i
; i
--) {
2578 nsh
= alloc_stripe(sc
, GFP_KERNEL
, newsize
, conf
);
2582 list_add(&nsh
->lru
, &newstripes
);
2585 /* didn't get enough, give up */
2586 while (!list_empty(&newstripes
)) {
2587 nsh
= list_entry(newstripes
.next
, struct stripe_head
, lru
);
2588 list_del(&nsh
->lru
);
2589 free_stripe(sc
, nsh
);
2591 kmem_cache_destroy(sc
);
2592 mutex_unlock(&conf
->cache_size_mutex
);
2595 /* Step 2 - Must use GFP_NOIO now.
2596 * OK, we have enough stripes, start collecting inactive
2597 * stripes and copying them over
2601 list_for_each_entry(nsh
, &newstripes
, lru
) {
2602 lock_device_hash_lock(conf
, hash
);
2603 wait_event_cmd(conf
->wait_for_stripe
,
2604 !list_empty(conf
->inactive_list
+ hash
),
2605 unlock_device_hash_lock(conf
, hash
),
2606 lock_device_hash_lock(conf
, hash
));
2607 osh
= get_free_stripe(conf
, hash
);
2608 unlock_device_hash_lock(conf
, hash
);
2610 #if PAGE_SIZE != DEFAULT_STRIPE_SIZE
2611 for (i
= 0; i
< osh
->nr_pages
; i
++) {
2612 nsh
->pages
[i
] = osh
->pages
[i
];
2613 osh
->pages
[i
] = NULL
;
2616 for(i
=0; i
<conf
->pool_size
; i
++) {
2617 nsh
->dev
[i
].page
= osh
->dev
[i
].page
;
2618 nsh
->dev
[i
].orig_page
= osh
->dev
[i
].page
;
2619 nsh
->dev
[i
].offset
= osh
->dev
[i
].offset
;
2621 nsh
->hash_lock_index
= hash
;
2622 free_stripe(conf
->slab_cache
, osh
);
2624 if (cnt
>= conf
->max_nr_stripes
/ NR_STRIPE_HASH_LOCKS
+
2625 !!((conf
->max_nr_stripes
% NR_STRIPE_HASH_LOCKS
) > hash
)) {
2630 kmem_cache_destroy(conf
->slab_cache
);
2633 * At this point, we are holding all the stripes so the array
2634 * is completely stalled, so now is a good time to resize
2635 * conf->disks and the scribble region
2637 ndisks
= kcalloc(newsize
, sizeof(struct disk_info
), GFP_NOIO
);
2639 for (i
= 0; i
< conf
->pool_size
; i
++)
2640 ndisks
[i
] = conf
->disks
[i
];
2642 for (i
= conf
->pool_size
; i
< newsize
; i
++) {
2643 ndisks
[i
].extra_page
= alloc_page(GFP_NOIO
);
2644 if (!ndisks
[i
].extra_page
)
2649 for (i
= conf
->pool_size
; i
< newsize
; i
++)
2650 if (ndisks
[i
].extra_page
)
2651 put_page(ndisks
[i
].extra_page
);
2655 conf
->disks
= ndisks
;
2660 conf
->slab_cache
= sc
;
2661 conf
->active_name
= 1-conf
->active_name
;
2663 /* Step 4, return new stripes to service */
2664 while(!list_empty(&newstripes
)) {
2665 nsh
= list_entry(newstripes
.next
, struct stripe_head
, lru
);
2666 list_del_init(&nsh
->lru
);
2668 #if PAGE_SIZE != DEFAULT_STRIPE_SIZE
2669 for (i
= 0; i
< nsh
->nr_pages
; i
++) {
2672 nsh
->pages
[i
] = alloc_page(GFP_NOIO
);
2677 for (i
= conf
->raid_disks
; i
< newsize
; i
++) {
2678 if (nsh
->dev
[i
].page
)
2680 nsh
->dev
[i
].page
= raid5_get_dev_page(nsh
, i
);
2681 nsh
->dev
[i
].orig_page
= nsh
->dev
[i
].page
;
2682 nsh
->dev
[i
].offset
= raid5_get_page_offset(nsh
, i
);
2685 for (i
=conf
->raid_disks
; i
< newsize
; i
++)
2686 if (nsh
->dev
[i
].page
== NULL
) {
2687 struct page
*p
= alloc_page(GFP_NOIO
);
2688 nsh
->dev
[i
].page
= p
;
2689 nsh
->dev
[i
].orig_page
= p
;
2690 nsh
->dev
[i
].offset
= 0;
2695 raid5_release_stripe(nsh
);
2697 /* critical section pass, GFP_NOIO no longer needed */
2700 conf
->pool_size
= newsize
;
2701 mutex_unlock(&conf
->cache_size_mutex
);
2706 static int drop_one_stripe(struct r5conf
*conf
)
2708 struct stripe_head
*sh
;
2709 int hash
= (conf
->max_nr_stripes
- 1) & STRIPE_HASH_LOCKS_MASK
;
2711 spin_lock_irq(conf
->hash_locks
+ hash
);
2712 sh
= get_free_stripe(conf
, hash
);
2713 spin_unlock_irq(conf
->hash_locks
+ hash
);
2716 BUG_ON(atomic_read(&sh
->count
));
2718 free_stripe(conf
->slab_cache
, sh
);
2719 atomic_dec(&conf
->active_stripes
);
2720 conf
->max_nr_stripes
--;
2724 static void shrink_stripes(struct r5conf
*conf
)
2726 while (conf
->max_nr_stripes
&&
2727 drop_one_stripe(conf
))
2730 kmem_cache_destroy(conf
->slab_cache
);
2731 conf
->slab_cache
= NULL
;
2735 * This helper wraps rcu_dereference_protected() and can be used when
2736 * it is known that the nr_pending of the rdev is elevated.
2738 static struct md_rdev
*rdev_pend_deref(struct md_rdev __rcu
*rdev
)
2740 return rcu_dereference_protected(rdev
,
2741 atomic_read(&rcu_access_pointer(rdev
)->nr_pending
));
2745 * This helper wraps rcu_dereference_protected() and should be used
2746 * when it is known that the mddev_lock() is held. This is safe
2747 * seeing raid5_remove_disk() has the same lock held.
2749 static struct md_rdev
*rdev_mdlock_deref(struct mddev
*mddev
,
2750 struct md_rdev __rcu
*rdev
)
2752 return rcu_dereference_protected(rdev
,
2753 lockdep_is_held(&mddev
->reconfig_mutex
));
2756 static void raid5_end_read_request(struct bio
* bi
)
2758 struct stripe_head
*sh
= bi
->bi_private
;
2759 struct r5conf
*conf
= sh
->raid_conf
;
2760 int disks
= sh
->disks
, i
;
2761 struct md_rdev
*rdev
= NULL
;
2764 for (i
=0 ; i
<disks
; i
++)
2765 if (bi
== &sh
->dev
[i
].req
)
2768 pr_debug("end_read_request %llu/%d, count: %d, error %d.\n",
2769 (unsigned long long)sh
->sector
, i
, atomic_read(&sh
->count
),
2775 if (test_bit(R5_ReadRepl
, &sh
->dev
[i
].flags
))
2776 /* If replacement finished while this request was outstanding,
2777 * 'replacement' might be NULL already.
2778 * In that case it moved down to 'rdev'.
2779 * rdev is not removed until all requests are finished.
2781 rdev
= rdev_pend_deref(conf
->disks
[i
].replacement
);
2783 rdev
= rdev_pend_deref(conf
->disks
[i
].rdev
);
2785 if (use_new_offset(conf
, sh
))
2786 s
= sh
->sector
+ rdev
->new_data_offset
;
2788 s
= sh
->sector
+ rdev
->data_offset
;
2789 if (!bi
->bi_status
) {
2790 set_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
);
2791 if (test_bit(R5_ReadError
, &sh
->dev
[i
].flags
)) {
2792 /* Note that this cannot happen on a
2793 * replacement device. We just fail those on
2796 pr_info_ratelimited(
2797 "md/raid:%s: read error corrected (%lu sectors at %llu on %pg)\n",
2798 mdname(conf
->mddev
), RAID5_STRIPE_SECTORS(conf
),
2799 (unsigned long long)s
,
2801 atomic_add(RAID5_STRIPE_SECTORS(conf
), &rdev
->corrected_errors
);
2802 clear_bit(R5_ReadError
, &sh
->dev
[i
].flags
);
2803 clear_bit(R5_ReWrite
, &sh
->dev
[i
].flags
);
2804 } else if (test_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
))
2805 clear_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
);
2807 if (test_bit(R5_InJournal
, &sh
->dev
[i
].flags
))
2809 * end read for a page in journal, this
2810 * must be preparing for prexor in rmw
2812 set_bit(R5_OrigPageUPTDODATE
, &sh
->dev
[i
].flags
);
2814 if (atomic_read(&rdev
->read_errors
))
2815 atomic_set(&rdev
->read_errors
, 0);
2820 clear_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
);
2821 if (!(bi
->bi_status
== BLK_STS_PROTECTION
))
2822 atomic_inc(&rdev
->read_errors
);
2823 if (test_bit(R5_ReadRepl
, &sh
->dev
[i
].flags
))
2824 pr_warn_ratelimited(
2825 "md/raid:%s: read error on replacement device (sector %llu on %pg).\n",
2826 mdname(conf
->mddev
),
2827 (unsigned long long)s
,
2829 else if (conf
->mddev
->degraded
>= conf
->max_degraded
) {
2831 pr_warn_ratelimited(
2832 "md/raid:%s: read error not correctable (sector %llu on %pg).\n",
2833 mdname(conf
->mddev
),
2834 (unsigned long long)s
,
2836 } else if (test_bit(R5_ReWrite
, &sh
->dev
[i
].flags
)) {
2839 pr_warn_ratelimited(
2840 "md/raid:%s: read error NOT corrected!! (sector %llu on %pg).\n",
2841 mdname(conf
->mddev
),
2842 (unsigned long long)s
,
2844 } else if (atomic_read(&rdev
->read_errors
)
2845 > conf
->max_nr_stripes
) {
2846 if (!test_bit(Faulty
, &rdev
->flags
)) {
2847 pr_warn("md/raid:%s: %d read_errors > %d stripes\n",
2848 mdname(conf
->mddev
),
2849 atomic_read(&rdev
->read_errors
),
2850 conf
->max_nr_stripes
);
2851 pr_warn("md/raid:%s: Too many read errors, failing device %pg.\n",
2852 mdname(conf
->mddev
), rdev
->bdev
);
2856 if (set_bad
&& test_bit(In_sync
, &rdev
->flags
)
2857 && !test_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
))
2860 if (sh
->qd_idx
>= 0 && sh
->pd_idx
== i
)
2861 set_bit(R5_ReadError
, &sh
->dev
[i
].flags
);
2862 else if (test_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
)) {
2863 set_bit(R5_ReadError
, &sh
->dev
[i
].flags
);
2864 clear_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
);
2866 set_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
);
2868 clear_bit(R5_ReadError
, &sh
->dev
[i
].flags
);
2869 clear_bit(R5_ReWrite
, &sh
->dev
[i
].flags
);
2871 && test_bit(In_sync
, &rdev
->flags
)
2872 && rdev_set_badblocks(
2873 rdev
, sh
->sector
, RAID5_STRIPE_SECTORS(conf
), 0)))
2874 md_error(conf
->mddev
, rdev
);
2877 rdev_dec_pending(rdev
, conf
->mddev
);
2879 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
2880 set_bit(STRIPE_HANDLE
, &sh
->state
);
2881 raid5_release_stripe(sh
);
2884 static void raid5_end_write_request(struct bio
*bi
)
2886 struct stripe_head
*sh
= bi
->bi_private
;
2887 struct r5conf
*conf
= sh
->raid_conf
;
2888 int disks
= sh
->disks
, i
;
2889 struct md_rdev
*rdev
;
2892 int replacement
= 0;
2894 for (i
= 0 ; i
< disks
; i
++) {
2895 if (bi
== &sh
->dev
[i
].req
) {
2896 rdev
= rdev_pend_deref(conf
->disks
[i
].rdev
);
2899 if (bi
== &sh
->dev
[i
].rreq
) {
2900 rdev
= rdev_pend_deref(conf
->disks
[i
].replacement
);
2904 /* rdev was removed and 'replacement'
2905 * replaced it. rdev is not removed
2906 * until all requests are finished.
2908 rdev
= rdev_pend_deref(conf
->disks
[i
].rdev
);
2912 pr_debug("end_write_request %llu/%d, count %d, error: %d.\n",
2913 (unsigned long long)sh
->sector
, i
, atomic_read(&sh
->count
),
2922 md_error(conf
->mddev
, rdev
);
2923 else if (is_badblock(rdev
, sh
->sector
,
2924 RAID5_STRIPE_SECTORS(conf
),
2925 &first_bad
, &bad_sectors
))
2926 set_bit(R5_MadeGoodRepl
, &sh
->dev
[i
].flags
);
2928 if (bi
->bi_status
) {
2929 set_bit(STRIPE_DEGRADED
, &sh
->state
);
2930 set_bit(WriteErrorSeen
, &rdev
->flags
);
2931 set_bit(R5_WriteError
, &sh
->dev
[i
].flags
);
2932 if (!test_and_set_bit(WantReplacement
, &rdev
->flags
))
2933 set_bit(MD_RECOVERY_NEEDED
,
2934 &rdev
->mddev
->recovery
);
2935 } else if (is_badblock(rdev
, sh
->sector
,
2936 RAID5_STRIPE_SECTORS(conf
),
2937 &first_bad
, &bad_sectors
)) {
2938 set_bit(R5_MadeGood
, &sh
->dev
[i
].flags
);
2939 if (test_bit(R5_ReadError
, &sh
->dev
[i
].flags
))
2940 /* That was a successful write so make
2941 * sure it looks like we already did
2944 set_bit(R5_ReWrite
, &sh
->dev
[i
].flags
);
2947 rdev_dec_pending(rdev
, conf
->mddev
);
2949 if (sh
->batch_head
&& bi
->bi_status
&& !replacement
)
2950 set_bit(STRIPE_BATCH_ERR
, &sh
->batch_head
->state
);
2953 if (!test_and_clear_bit(R5_DOUBLE_LOCKED
, &sh
->dev
[i
].flags
))
2954 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
2955 set_bit(STRIPE_HANDLE
, &sh
->state
);
2957 if (sh
->batch_head
&& sh
!= sh
->batch_head
)
2958 raid5_release_stripe(sh
->batch_head
);
2959 raid5_release_stripe(sh
);
2962 static void raid5_error(struct mddev
*mddev
, struct md_rdev
*rdev
)
2964 struct r5conf
*conf
= mddev
->private;
2965 unsigned long flags
;
2966 pr_debug("raid456: error called\n");
2968 pr_crit("md/raid:%s: Disk failure on %pg, disabling device.\n",
2969 mdname(mddev
), rdev
->bdev
);
2971 spin_lock_irqsave(&conf
->device_lock
, flags
);
2972 set_bit(Faulty
, &rdev
->flags
);
2973 clear_bit(In_sync
, &rdev
->flags
);
2974 mddev
->degraded
= raid5_calc_degraded(conf
);
2976 if (has_failed(conf
)) {
2977 set_bit(MD_BROKEN
, &conf
->mddev
->flags
);
2978 conf
->recovery_disabled
= mddev
->recovery_disabled
;
2980 pr_crit("md/raid:%s: Cannot continue operation (%d/%d failed).\n",
2981 mdname(mddev
), mddev
->degraded
, conf
->raid_disks
);
2983 pr_crit("md/raid:%s: Operation continuing on %d devices.\n",
2984 mdname(mddev
), conf
->raid_disks
- mddev
->degraded
);
2987 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
2988 set_bit(MD_RECOVERY_INTR
, &mddev
->recovery
);
2990 set_bit(Blocked
, &rdev
->flags
);
2991 set_mask_bits(&mddev
->sb_flags
, 0,
2992 BIT(MD_SB_CHANGE_DEVS
) | BIT(MD_SB_CHANGE_PENDING
));
2993 r5c_update_on_rdev_error(mddev
, rdev
);
2997 * Input: a 'big' sector number,
2998 * Output: index of the data and parity disk, and the sector # in them.
3000 sector_t
raid5_compute_sector(struct r5conf
*conf
, sector_t r_sector
,
3001 int previous
, int *dd_idx
,
3002 struct stripe_head
*sh
)
3004 sector_t stripe
, stripe2
;
3005 sector_t chunk_number
;
3006 unsigned int chunk_offset
;
3009 sector_t new_sector
;
3010 int algorithm
= previous
? conf
->prev_algo
3012 int sectors_per_chunk
= previous
? conf
->prev_chunk_sectors
3013 : conf
->chunk_sectors
;
3014 int raid_disks
= previous
? conf
->previous_raid_disks
3016 int data_disks
= raid_disks
- conf
->max_degraded
;
3018 /* First compute the information on this sector */
3021 * Compute the chunk number and the sector offset inside the chunk
3023 chunk_offset
= sector_div(r_sector
, sectors_per_chunk
);
3024 chunk_number
= r_sector
;
3027 * Compute the stripe number
3029 stripe
= chunk_number
;
3030 *dd_idx
= sector_div(stripe
, data_disks
);
3033 * Select the parity disk based on the user selected algorithm.
3035 pd_idx
= qd_idx
= -1;
3036 switch(conf
->level
) {
3038 pd_idx
= data_disks
;
3041 switch (algorithm
) {
3042 case ALGORITHM_LEFT_ASYMMETRIC
:
3043 pd_idx
= data_disks
- sector_div(stripe2
, raid_disks
);
3044 if (*dd_idx
>= pd_idx
)
3047 case ALGORITHM_RIGHT_ASYMMETRIC
:
3048 pd_idx
= sector_div(stripe2
, raid_disks
);
3049 if (*dd_idx
>= pd_idx
)
3052 case ALGORITHM_LEFT_SYMMETRIC
:
3053 pd_idx
= data_disks
- sector_div(stripe2
, raid_disks
);
3054 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % raid_disks
;
3056 case ALGORITHM_RIGHT_SYMMETRIC
:
3057 pd_idx
= sector_div(stripe2
, raid_disks
);
3058 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % raid_disks
;
3060 case ALGORITHM_PARITY_0
:
3064 case ALGORITHM_PARITY_N
:
3065 pd_idx
= data_disks
;
3073 switch (algorithm
) {
3074 case ALGORITHM_LEFT_ASYMMETRIC
:
3075 pd_idx
= raid_disks
- 1 - sector_div(stripe2
, raid_disks
);
3076 qd_idx
= pd_idx
+ 1;
3077 if (pd_idx
== raid_disks
-1) {
3078 (*dd_idx
)++; /* Q D D D P */
3080 } else if (*dd_idx
>= pd_idx
)
3081 (*dd_idx
) += 2; /* D D P Q D */
3083 case ALGORITHM_RIGHT_ASYMMETRIC
:
3084 pd_idx
= sector_div(stripe2
, raid_disks
);
3085 qd_idx
= pd_idx
+ 1;
3086 if (pd_idx
== raid_disks
-1) {
3087 (*dd_idx
)++; /* Q D D D P */
3089 } else if (*dd_idx
>= pd_idx
)
3090 (*dd_idx
) += 2; /* D D P Q D */
3092 case ALGORITHM_LEFT_SYMMETRIC
:
3093 pd_idx
= raid_disks
- 1 - sector_div(stripe2
, raid_disks
);
3094 qd_idx
= (pd_idx
+ 1) % raid_disks
;
3095 *dd_idx
= (pd_idx
+ 2 + *dd_idx
) % raid_disks
;
3097 case ALGORITHM_RIGHT_SYMMETRIC
:
3098 pd_idx
= sector_div(stripe2
, raid_disks
);
3099 qd_idx
= (pd_idx
+ 1) % raid_disks
;
3100 *dd_idx
= (pd_idx
+ 2 + *dd_idx
) % raid_disks
;
3103 case ALGORITHM_PARITY_0
:
3108 case ALGORITHM_PARITY_N
:
3109 pd_idx
= data_disks
;
3110 qd_idx
= data_disks
+ 1;
3113 case ALGORITHM_ROTATING_ZERO_RESTART
:
3114 /* Exactly the same as RIGHT_ASYMMETRIC, but or
3115 * of blocks for computing Q is different.
3117 pd_idx
= sector_div(stripe2
, raid_disks
);
3118 qd_idx
= pd_idx
+ 1;
3119 if (pd_idx
== raid_disks
-1) {
3120 (*dd_idx
)++; /* Q D D D P */
3122 } else if (*dd_idx
>= pd_idx
)
3123 (*dd_idx
) += 2; /* D D P Q D */
3127 case ALGORITHM_ROTATING_N_RESTART
:
3128 /* Same a left_asymmetric, by first stripe is
3129 * D D D P Q rather than
3133 pd_idx
= raid_disks
- 1 - sector_div(stripe2
, raid_disks
);
3134 qd_idx
= pd_idx
+ 1;
3135 if (pd_idx
== raid_disks
-1) {
3136 (*dd_idx
)++; /* Q D D D P */
3138 } else if (*dd_idx
>= pd_idx
)
3139 (*dd_idx
) += 2; /* D D P Q D */
3143 case ALGORITHM_ROTATING_N_CONTINUE
:
3144 /* Same as left_symmetric but Q is before P */
3145 pd_idx
= raid_disks
- 1 - sector_div(stripe2
, raid_disks
);
3146 qd_idx
= (pd_idx
+ raid_disks
- 1) % raid_disks
;
3147 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % raid_disks
;
3151 case ALGORITHM_LEFT_ASYMMETRIC_6
:
3152 /* RAID5 left_asymmetric, with Q on last device */
3153 pd_idx
= data_disks
- sector_div(stripe2
, raid_disks
-1);
3154 if (*dd_idx
>= pd_idx
)
3156 qd_idx
= raid_disks
- 1;
3159 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
3160 pd_idx
= sector_div(stripe2
, raid_disks
-1);
3161 if (*dd_idx
>= pd_idx
)
3163 qd_idx
= raid_disks
- 1;
3166 case ALGORITHM_LEFT_SYMMETRIC_6
:
3167 pd_idx
= data_disks
- sector_div(stripe2
, raid_disks
-1);
3168 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % (raid_disks
-1);
3169 qd_idx
= raid_disks
- 1;
3172 case ALGORITHM_RIGHT_SYMMETRIC_6
:
3173 pd_idx
= sector_div(stripe2
, raid_disks
-1);
3174 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % (raid_disks
-1);
3175 qd_idx
= raid_disks
- 1;
3178 case ALGORITHM_PARITY_0_6
:
3181 qd_idx
= raid_disks
- 1;
3191 sh
->pd_idx
= pd_idx
;
3192 sh
->qd_idx
= qd_idx
;
3193 sh
->ddf_layout
= ddf_layout
;
3196 * Finally, compute the new sector number
3198 new_sector
= (sector_t
)stripe
* sectors_per_chunk
+ chunk_offset
;
3202 sector_t
raid5_compute_blocknr(struct stripe_head
*sh
, int i
, int previous
)
3204 struct r5conf
*conf
= sh
->raid_conf
;
3205 int raid_disks
= sh
->disks
;
3206 int data_disks
= raid_disks
- conf
->max_degraded
;
3207 sector_t new_sector
= sh
->sector
, check
;
3208 int sectors_per_chunk
= previous
? conf
->prev_chunk_sectors
3209 : conf
->chunk_sectors
;
3210 int algorithm
= previous
? conf
->prev_algo
3214 sector_t chunk_number
;
3215 int dummy1
, dd_idx
= i
;
3217 struct stripe_head sh2
;
3219 chunk_offset
= sector_div(new_sector
, sectors_per_chunk
);
3220 stripe
= new_sector
;
3222 if (i
== sh
->pd_idx
)
3224 switch(conf
->level
) {
3227 switch (algorithm
) {
3228 case ALGORITHM_LEFT_ASYMMETRIC
:
3229 case ALGORITHM_RIGHT_ASYMMETRIC
:
3233 case ALGORITHM_LEFT_SYMMETRIC
:
3234 case ALGORITHM_RIGHT_SYMMETRIC
:
3237 i
-= (sh
->pd_idx
+ 1);
3239 case ALGORITHM_PARITY_0
:
3242 case ALGORITHM_PARITY_N
:
3249 if (i
== sh
->qd_idx
)
3250 return 0; /* It is the Q disk */
3251 switch (algorithm
) {
3252 case ALGORITHM_LEFT_ASYMMETRIC
:
3253 case ALGORITHM_RIGHT_ASYMMETRIC
:
3254 case ALGORITHM_ROTATING_ZERO_RESTART
:
3255 case ALGORITHM_ROTATING_N_RESTART
:
3256 if (sh
->pd_idx
== raid_disks
-1)
3257 i
--; /* Q D D D P */
3258 else if (i
> sh
->pd_idx
)
3259 i
-= 2; /* D D P Q D */
3261 case ALGORITHM_LEFT_SYMMETRIC
:
3262 case ALGORITHM_RIGHT_SYMMETRIC
:
3263 if (sh
->pd_idx
== raid_disks
-1)
3264 i
--; /* Q D D D P */
3269 i
-= (sh
->pd_idx
+ 2);
3272 case ALGORITHM_PARITY_0
:
3275 case ALGORITHM_PARITY_N
:
3277 case ALGORITHM_ROTATING_N_CONTINUE
:
3278 /* Like left_symmetric, but P is before Q */
3279 if (sh
->pd_idx
== 0)
3280 i
--; /* P D D D Q */
3285 i
-= (sh
->pd_idx
+ 1);
3288 case ALGORITHM_LEFT_ASYMMETRIC_6
:
3289 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
3293 case ALGORITHM_LEFT_SYMMETRIC_6
:
3294 case ALGORITHM_RIGHT_SYMMETRIC_6
:
3296 i
+= data_disks
+ 1;
3297 i
-= (sh
->pd_idx
+ 1);
3299 case ALGORITHM_PARITY_0_6
:
3308 chunk_number
= stripe
* data_disks
+ i
;
3309 r_sector
= chunk_number
* sectors_per_chunk
+ chunk_offset
;
3311 check
= raid5_compute_sector(conf
, r_sector
,
3312 previous
, &dummy1
, &sh2
);
3313 if (check
!= sh
->sector
|| dummy1
!= dd_idx
|| sh2
.pd_idx
!= sh
->pd_idx
3314 || sh2
.qd_idx
!= sh
->qd_idx
) {
3315 pr_warn("md/raid:%s: compute_blocknr: map not correct\n",
3316 mdname(conf
->mddev
));
3323 * There are cases where we want handle_stripe_dirtying() and
3324 * schedule_reconstruction() to delay towrite to some dev of a stripe.
3326 * This function checks whether we want to delay the towrite. Specifically,
3327 * we delay the towrite when:
3329 * 1. degraded stripe has a non-overwrite to the missing dev, AND this
3330 * stripe has data in journal (for other devices).
3332 * In this case, when reading data for the non-overwrite dev, it is
3333 * necessary to handle complex rmw of write back cache (prexor with
3334 * orig_page, and xor with page). To keep read path simple, we would
3335 * like to flush data in journal to RAID disks first, so complex rmw
3336 * is handled in the write patch (handle_stripe_dirtying).
3338 * 2. when journal space is critical (R5C_LOG_CRITICAL=1)
3340 * It is important to be able to flush all stripes in raid5-cache.
3341 * Therefore, we need reserve some space on the journal device for
3342 * these flushes. If flush operation includes pending writes to the
3343 * stripe, we need to reserve (conf->raid_disk + 1) pages per stripe
3344 * for the flush out. If we exclude these pending writes from flush
3345 * operation, we only need (conf->max_degraded + 1) pages per stripe.
3346 * Therefore, excluding pending writes in these cases enables more
3347 * efficient use of the journal device.
3349 * Note: To make sure the stripe makes progress, we only delay
3350 * towrite for stripes with data already in journal (injournal > 0).
3351 * When LOG_CRITICAL, stripes with injournal == 0 will be sent to
3352 * no_space_stripes list.
3354 * 3. during journal failure
3355 * In journal failure, we try to flush all cached data to raid disks
3356 * based on data in stripe cache. The array is read-only to upper
3357 * layers, so we would skip all pending writes.
3360 static inline bool delay_towrite(struct r5conf
*conf
,
3362 struct stripe_head_state
*s
)
3365 if (!test_bit(R5_OVERWRITE
, &dev
->flags
) &&
3366 !test_bit(R5_Insync
, &dev
->flags
) && s
->injournal
)
3369 if (test_bit(R5C_LOG_CRITICAL
, &conf
->cache_state
) &&
3373 if (s
->log_failed
&& s
->injournal
)
3379 schedule_reconstruction(struct stripe_head
*sh
, struct stripe_head_state
*s
,
3380 int rcw
, int expand
)
3382 int i
, pd_idx
= sh
->pd_idx
, qd_idx
= sh
->qd_idx
, disks
= sh
->disks
;
3383 struct r5conf
*conf
= sh
->raid_conf
;
3384 int level
= conf
->level
;
3388 * In some cases, handle_stripe_dirtying initially decided to
3389 * run rmw and allocates extra page for prexor. However, rcw is
3390 * cheaper later on. We need to free the extra page now,
3391 * because we won't be able to do that in ops_complete_prexor().
3393 r5c_release_extra_page(sh
);
3395 for (i
= disks
; i
--; ) {
3396 struct r5dev
*dev
= &sh
->dev
[i
];
3398 if (dev
->towrite
&& !delay_towrite(conf
, dev
, s
)) {
3399 set_bit(R5_LOCKED
, &dev
->flags
);
3400 set_bit(R5_Wantdrain
, &dev
->flags
);
3402 clear_bit(R5_UPTODATE
, &dev
->flags
);
3404 } else if (test_bit(R5_InJournal
, &dev
->flags
)) {
3405 set_bit(R5_LOCKED
, &dev
->flags
);
3409 /* if we are not expanding this is a proper write request, and
3410 * there will be bios with new data to be drained into the
3415 /* False alarm, nothing to do */
3417 sh
->reconstruct_state
= reconstruct_state_drain_run
;
3418 set_bit(STRIPE_OP_BIODRAIN
, &s
->ops_request
);
3420 sh
->reconstruct_state
= reconstruct_state_run
;
3422 set_bit(STRIPE_OP_RECONSTRUCT
, &s
->ops_request
);
3424 if (s
->locked
+ conf
->max_degraded
== disks
)
3425 if (!test_and_set_bit(STRIPE_FULL_WRITE
, &sh
->state
))
3426 atomic_inc(&conf
->pending_full_writes
);
3428 BUG_ON(!(test_bit(R5_UPTODATE
, &sh
->dev
[pd_idx
].flags
) ||
3429 test_bit(R5_Wantcompute
, &sh
->dev
[pd_idx
].flags
)));
3430 BUG_ON(level
== 6 &&
3431 (!(test_bit(R5_UPTODATE
, &sh
->dev
[qd_idx
].flags
) ||
3432 test_bit(R5_Wantcompute
, &sh
->dev
[qd_idx
].flags
))));
3434 for (i
= disks
; i
--; ) {
3435 struct r5dev
*dev
= &sh
->dev
[i
];
3436 if (i
== pd_idx
|| i
== qd_idx
)
3440 (test_bit(R5_UPTODATE
, &dev
->flags
) ||
3441 test_bit(R5_Wantcompute
, &dev
->flags
))) {
3442 set_bit(R5_Wantdrain
, &dev
->flags
);
3443 set_bit(R5_LOCKED
, &dev
->flags
);
3444 clear_bit(R5_UPTODATE
, &dev
->flags
);
3446 } else if (test_bit(R5_InJournal
, &dev
->flags
)) {
3447 set_bit(R5_LOCKED
, &dev
->flags
);
3452 /* False alarm - nothing to do */
3454 sh
->reconstruct_state
= reconstruct_state_prexor_drain_run
;
3455 set_bit(STRIPE_OP_PREXOR
, &s
->ops_request
);
3456 set_bit(STRIPE_OP_BIODRAIN
, &s
->ops_request
);
3457 set_bit(STRIPE_OP_RECONSTRUCT
, &s
->ops_request
);
3460 /* keep the parity disk(s) locked while asynchronous operations
3463 set_bit(R5_LOCKED
, &sh
->dev
[pd_idx
].flags
);
3464 clear_bit(R5_UPTODATE
, &sh
->dev
[pd_idx
].flags
);
3468 int qd_idx
= sh
->qd_idx
;
3469 struct r5dev
*dev
= &sh
->dev
[qd_idx
];
3471 set_bit(R5_LOCKED
, &dev
->flags
);
3472 clear_bit(R5_UPTODATE
, &dev
->flags
);
3476 if (raid5_has_ppl(sh
->raid_conf
) && sh
->ppl_page
&&
3477 test_bit(STRIPE_OP_BIODRAIN
, &s
->ops_request
) &&
3478 !test_bit(STRIPE_FULL_WRITE
, &sh
->state
) &&
3479 test_bit(R5_Insync
, &sh
->dev
[pd_idx
].flags
))
3480 set_bit(STRIPE_OP_PARTIAL_PARITY
, &s
->ops_request
);
3482 pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
3483 __func__
, (unsigned long long)sh
->sector
,
3484 s
->locked
, s
->ops_request
);
3487 static bool stripe_bio_overlaps(struct stripe_head
*sh
, struct bio
*bi
,
3488 int dd_idx
, int forwrite
)
3490 struct r5conf
*conf
= sh
->raid_conf
;
3493 pr_debug("checking bi b#%llu to stripe s#%llu\n",
3494 bi
->bi_iter
.bi_sector
, sh
->sector
);
3496 /* Don't allow new IO added to stripes in batch list */
3501 bip
= &sh
->dev
[dd_idx
].towrite
;
3503 bip
= &sh
->dev
[dd_idx
].toread
;
3505 while (*bip
&& (*bip
)->bi_iter
.bi_sector
< bi
->bi_iter
.bi_sector
) {
3506 if (bio_end_sector(*bip
) > bi
->bi_iter
.bi_sector
)
3508 bip
= &(*bip
)->bi_next
;
3511 if (*bip
&& (*bip
)->bi_iter
.bi_sector
< bio_end_sector(bi
))
3514 if (forwrite
&& raid5_has_ppl(conf
)) {
3516 * With PPL only writes to consecutive data chunks within a
3517 * stripe are allowed because for a single stripe_head we can
3518 * only have one PPL entry at a time, which describes one data
3519 * range. Not really an overlap, but wait_for_overlap can be
3520 * used to handle this.
3528 for (i
= 0; i
< sh
->disks
; i
++) {
3529 if (i
!= sh
->pd_idx
&&
3530 (i
== dd_idx
|| sh
->dev
[i
].towrite
)) {
3531 sector
= sh
->dev
[i
].sector
;
3532 if (count
== 0 || sector
< first
)
3540 if (first
+ conf
->chunk_sectors
* (count
- 1) != last
)
3547 static void __add_stripe_bio(struct stripe_head
*sh
, struct bio
*bi
,
3548 int dd_idx
, int forwrite
, int previous
)
3550 struct r5conf
*conf
= sh
->raid_conf
;
3555 bip
= &sh
->dev
[dd_idx
].towrite
;
3559 bip
= &sh
->dev
[dd_idx
].toread
;
3562 while (*bip
&& (*bip
)->bi_iter
.bi_sector
< bi
->bi_iter
.bi_sector
)
3563 bip
= &(*bip
)->bi_next
;
3565 if (!forwrite
|| previous
)
3566 clear_bit(STRIPE_BATCH_READY
, &sh
->state
);
3568 BUG_ON(*bip
&& bi
->bi_next
&& (*bip
) != bi
->bi_next
);
3572 bio_inc_remaining(bi
);
3573 md_write_inc(conf
->mddev
, bi
);
3576 /* check if page is covered */
3577 sector_t sector
= sh
->dev
[dd_idx
].sector
;
3578 for (bi
=sh
->dev
[dd_idx
].towrite
;
3579 sector
< sh
->dev
[dd_idx
].sector
+ RAID5_STRIPE_SECTORS(conf
) &&
3580 bi
&& bi
->bi_iter
.bi_sector
<= sector
;
3581 bi
= r5_next_bio(conf
, bi
, sh
->dev
[dd_idx
].sector
)) {
3582 if (bio_end_sector(bi
) >= sector
)
3583 sector
= bio_end_sector(bi
);
3585 if (sector
>= sh
->dev
[dd_idx
].sector
+ RAID5_STRIPE_SECTORS(conf
))
3586 if (!test_and_set_bit(R5_OVERWRITE
, &sh
->dev
[dd_idx
].flags
))
3587 sh
->overwrite_disks
++;
3590 pr_debug("added bi b#%llu to stripe s#%llu, disk %d, logical %llu\n",
3591 (*bip
)->bi_iter
.bi_sector
, sh
->sector
, dd_idx
,
3592 sh
->dev
[dd_idx
].sector
);
3594 if (conf
->mddev
->bitmap
&& firstwrite
) {
3595 /* Cannot hold spinlock over bitmap_startwrite,
3596 * but must ensure this isn't added to a batch until
3597 * we have added to the bitmap and set bm_seq.
3598 * So set STRIPE_BITMAP_PENDING to prevent
3600 * If multiple __add_stripe_bio() calls race here they
3601 * much all set STRIPE_BITMAP_PENDING. So only the first one
3602 * to complete "bitmap_startwrite" gets to set
3603 * STRIPE_BIT_DELAY. This is important as once a stripe
3604 * is added to a batch, STRIPE_BIT_DELAY cannot be changed
3607 set_bit(STRIPE_BITMAP_PENDING
, &sh
->state
);
3608 spin_unlock_irq(&sh
->stripe_lock
);
3609 md_bitmap_startwrite(conf
->mddev
->bitmap
, sh
->sector
,
3610 RAID5_STRIPE_SECTORS(conf
), 0);
3611 spin_lock_irq(&sh
->stripe_lock
);
3612 clear_bit(STRIPE_BITMAP_PENDING
, &sh
->state
);
3613 if (!sh
->batch_head
) {
3614 sh
->bm_seq
= conf
->seq_flush
+1;
3615 set_bit(STRIPE_BIT_DELAY
, &sh
->state
);
3621 * Each stripe/dev can have one or more bios attached.
3622 * toread/towrite point to the first in a chain.
3623 * The bi_next chain must be in order.
3625 static bool add_stripe_bio(struct stripe_head
*sh
, struct bio
*bi
,
3626 int dd_idx
, int forwrite
, int previous
)
3628 spin_lock_irq(&sh
->stripe_lock
);
3630 if (stripe_bio_overlaps(sh
, bi
, dd_idx
, forwrite
)) {
3631 set_bit(R5_Overlap
, &sh
->dev
[dd_idx
].flags
);
3632 spin_unlock_irq(&sh
->stripe_lock
);
3636 __add_stripe_bio(sh
, bi
, dd_idx
, forwrite
, previous
);
3637 spin_unlock_irq(&sh
->stripe_lock
);
3641 static void end_reshape(struct r5conf
*conf
);
3643 static void stripe_set_idx(sector_t stripe
, struct r5conf
*conf
, int previous
,
3644 struct stripe_head
*sh
)
3646 int sectors_per_chunk
=
3647 previous
? conf
->prev_chunk_sectors
: conf
->chunk_sectors
;
3649 int chunk_offset
= sector_div(stripe
, sectors_per_chunk
);
3650 int disks
= previous
? conf
->previous_raid_disks
: conf
->raid_disks
;
3652 raid5_compute_sector(conf
,
3653 stripe
* (disks
- conf
->max_degraded
)
3654 *sectors_per_chunk
+ chunk_offset
,
3660 handle_failed_stripe(struct r5conf
*conf
, struct stripe_head
*sh
,
3661 struct stripe_head_state
*s
, int disks
)
3664 BUG_ON(sh
->batch_head
);
3665 for (i
= disks
; i
--; ) {
3669 if (test_bit(R5_ReadError
, &sh
->dev
[i
].flags
)) {
3670 struct md_rdev
*rdev
;
3672 rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
3673 if (rdev
&& test_bit(In_sync
, &rdev
->flags
) &&
3674 !test_bit(Faulty
, &rdev
->flags
))
3675 atomic_inc(&rdev
->nr_pending
);
3680 if (!rdev_set_badblocks(
3683 RAID5_STRIPE_SECTORS(conf
), 0))
3684 md_error(conf
->mddev
, rdev
);
3685 rdev_dec_pending(rdev
, conf
->mddev
);
3688 spin_lock_irq(&sh
->stripe_lock
);
3689 /* fail all writes first */
3690 bi
= sh
->dev
[i
].towrite
;
3691 sh
->dev
[i
].towrite
= NULL
;
3692 sh
->overwrite_disks
= 0;
3693 spin_unlock_irq(&sh
->stripe_lock
);
3697 log_stripe_write_finished(sh
);
3699 if (test_and_clear_bit(R5_Overlap
, &sh
->dev
[i
].flags
))
3700 wake_up(&conf
->wait_for_overlap
);
3702 while (bi
&& bi
->bi_iter
.bi_sector
<
3703 sh
->dev
[i
].sector
+ RAID5_STRIPE_SECTORS(conf
)) {
3704 struct bio
*nextbi
= r5_next_bio(conf
, bi
, sh
->dev
[i
].sector
);
3706 md_write_end(conf
->mddev
);
3711 md_bitmap_endwrite(conf
->mddev
->bitmap
, sh
->sector
,
3712 RAID5_STRIPE_SECTORS(conf
), 0, 0);
3714 /* and fail all 'written' */
3715 bi
= sh
->dev
[i
].written
;
3716 sh
->dev
[i
].written
= NULL
;
3717 if (test_and_clear_bit(R5_SkipCopy
, &sh
->dev
[i
].flags
)) {
3718 WARN_ON(test_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
));
3719 sh
->dev
[i
].page
= sh
->dev
[i
].orig_page
;
3722 if (bi
) bitmap_end
= 1;
3723 while (bi
&& bi
->bi_iter
.bi_sector
<
3724 sh
->dev
[i
].sector
+ RAID5_STRIPE_SECTORS(conf
)) {
3725 struct bio
*bi2
= r5_next_bio(conf
, bi
, sh
->dev
[i
].sector
);
3727 md_write_end(conf
->mddev
);
3732 /* fail any reads if this device is non-operational and
3733 * the data has not reached the cache yet.
3735 if (!test_bit(R5_Wantfill
, &sh
->dev
[i
].flags
) &&
3736 s
->failed
> conf
->max_degraded
&&
3737 (!test_bit(R5_Insync
, &sh
->dev
[i
].flags
) ||
3738 test_bit(R5_ReadError
, &sh
->dev
[i
].flags
))) {
3739 spin_lock_irq(&sh
->stripe_lock
);
3740 bi
= sh
->dev
[i
].toread
;
3741 sh
->dev
[i
].toread
= NULL
;
3742 spin_unlock_irq(&sh
->stripe_lock
);
3743 if (test_and_clear_bit(R5_Overlap
, &sh
->dev
[i
].flags
))
3744 wake_up(&conf
->wait_for_overlap
);
3747 while (bi
&& bi
->bi_iter
.bi_sector
<
3748 sh
->dev
[i
].sector
+ RAID5_STRIPE_SECTORS(conf
)) {
3749 struct bio
*nextbi
=
3750 r5_next_bio(conf
, bi
, sh
->dev
[i
].sector
);
3757 md_bitmap_endwrite(conf
->mddev
->bitmap
, sh
->sector
,
3758 RAID5_STRIPE_SECTORS(conf
), 0, 0);
3759 /* If we were in the middle of a write the parity block might
3760 * still be locked - so just clear all R5_LOCKED flags
3762 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
3767 if (test_and_clear_bit(STRIPE_FULL_WRITE
, &sh
->state
))
3768 if (atomic_dec_and_test(&conf
->pending_full_writes
))
3769 md_wakeup_thread(conf
->mddev
->thread
);
3773 handle_failed_sync(struct r5conf
*conf
, struct stripe_head
*sh
,
3774 struct stripe_head_state
*s
)
3779 BUG_ON(sh
->batch_head
);
3780 clear_bit(STRIPE_SYNCING
, &sh
->state
);
3781 if (test_and_clear_bit(R5_Overlap
, &sh
->dev
[sh
->pd_idx
].flags
))
3782 wake_up(&conf
->wait_for_overlap
);
3785 /* There is nothing more to do for sync/check/repair.
3786 * Don't even need to abort as that is handled elsewhere
3787 * if needed, and not always wanted e.g. if there is a known
3789 * For recover/replace we need to record a bad block on all
3790 * non-sync devices, or abort the recovery
3792 if (test_bit(MD_RECOVERY_RECOVER
, &conf
->mddev
->recovery
)) {
3793 /* During recovery devices cannot be removed, so
3794 * locking and refcounting of rdevs is not needed
3797 for (i
= 0; i
< conf
->raid_disks
; i
++) {
3798 struct md_rdev
*rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
3800 && !test_bit(Faulty
, &rdev
->flags
)
3801 && !test_bit(In_sync
, &rdev
->flags
)
3802 && !rdev_set_badblocks(rdev
, sh
->sector
,
3803 RAID5_STRIPE_SECTORS(conf
), 0))
3805 rdev
= rcu_dereference(conf
->disks
[i
].replacement
);
3807 && !test_bit(Faulty
, &rdev
->flags
)
3808 && !test_bit(In_sync
, &rdev
->flags
)
3809 && !rdev_set_badblocks(rdev
, sh
->sector
,
3810 RAID5_STRIPE_SECTORS(conf
), 0))
3815 conf
->recovery_disabled
=
3816 conf
->mddev
->recovery_disabled
;
3818 md_done_sync(conf
->mddev
, RAID5_STRIPE_SECTORS(conf
), !abort
);
3821 static int want_replace(struct stripe_head
*sh
, int disk_idx
)
3823 struct md_rdev
*rdev
;
3827 rdev
= rcu_dereference(sh
->raid_conf
->disks
[disk_idx
].replacement
);
3829 && !test_bit(Faulty
, &rdev
->flags
)
3830 && !test_bit(In_sync
, &rdev
->flags
)
3831 && (rdev
->recovery_offset
<= sh
->sector
3832 || rdev
->mddev
->recovery_cp
<= sh
->sector
))
3838 static int need_this_block(struct stripe_head
*sh
, struct stripe_head_state
*s
,
3839 int disk_idx
, int disks
)
3841 struct r5dev
*dev
= &sh
->dev
[disk_idx
];
3842 struct r5dev
*fdev
[2] = { &sh
->dev
[s
->failed_num
[0]],
3843 &sh
->dev
[s
->failed_num
[1]] };
3845 bool force_rcw
= (sh
->raid_conf
->rmw_level
== PARITY_DISABLE_RMW
);
3848 if (test_bit(R5_LOCKED
, &dev
->flags
) ||
3849 test_bit(R5_UPTODATE
, &dev
->flags
))
3850 /* No point reading this as we already have it or have
3851 * decided to get it.
3856 (dev
->towrite
&& !test_bit(R5_OVERWRITE
, &dev
->flags
)))
3857 /* We need this block to directly satisfy a request */
3860 if (s
->syncing
|| s
->expanding
||
3861 (s
->replacing
&& want_replace(sh
, disk_idx
)))
3862 /* When syncing, or expanding we read everything.
3863 * When replacing, we need the replaced block.
3867 if ((s
->failed
>= 1 && fdev
[0]->toread
) ||
3868 (s
->failed
>= 2 && fdev
[1]->toread
))
3869 /* If we want to read from a failed device, then
3870 * we need to actually read every other device.
3874 /* Sometimes neither read-modify-write nor reconstruct-write
3875 * cycles can work. In those cases we read every block we
3876 * can. Then the parity-update is certain to have enough to
3878 * This can only be a problem when we need to write something,
3879 * and some device has failed. If either of those tests
3880 * fail we need look no further.
3882 if (!s
->failed
|| !s
->to_write
)
3885 if (test_bit(R5_Insync
, &dev
->flags
) &&
3886 !test_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
3887 /* Pre-reads at not permitted until after short delay
3888 * to gather multiple requests. However if this
3889 * device is no Insync, the block could only be computed
3890 * and there is no need to delay that.
3894 for (i
= 0; i
< s
->failed
&& i
< 2; i
++) {
3895 if (fdev
[i
]->towrite
&&
3896 !test_bit(R5_UPTODATE
, &fdev
[i
]->flags
) &&
3897 !test_bit(R5_OVERWRITE
, &fdev
[i
]->flags
))
3898 /* If we have a partial write to a failed
3899 * device, then we will need to reconstruct
3900 * the content of that device, so all other
3901 * devices must be read.
3905 if (s
->failed
>= 2 &&
3906 (fdev
[i
]->towrite
||
3907 s
->failed_num
[i
] == sh
->pd_idx
||
3908 s
->failed_num
[i
] == sh
->qd_idx
) &&
3909 !test_bit(R5_UPTODATE
, &fdev
[i
]->flags
))
3910 /* In max degraded raid6, If the failed disk is P, Q,
3911 * or we want to read the failed disk, we need to do
3912 * reconstruct-write.
3917 /* If we are forced to do a reconstruct-write, because parity
3918 * cannot be trusted and we are currently recovering it, there
3919 * is extra need to be careful.
3920 * If one of the devices that we would need to read, because
3921 * it is not being overwritten (and maybe not written at all)
3922 * is missing/faulty, then we need to read everything we can.
3925 sh
->sector
< sh
->raid_conf
->mddev
->recovery_cp
)
3926 /* reconstruct-write isn't being forced */
3928 for (i
= 0; i
< s
->failed
&& i
< 2; i
++) {
3929 if (s
->failed_num
[i
] != sh
->pd_idx
&&
3930 s
->failed_num
[i
] != sh
->qd_idx
&&
3931 !test_bit(R5_UPTODATE
, &fdev
[i
]->flags
) &&
3932 !test_bit(R5_OVERWRITE
, &fdev
[i
]->flags
))
3939 /* fetch_block - checks the given member device to see if its data needs
3940 * to be read or computed to satisfy a request.
3942 * Returns 1 when no more member devices need to be checked, otherwise returns
3943 * 0 to tell the loop in handle_stripe_fill to continue
3945 static int fetch_block(struct stripe_head
*sh
, struct stripe_head_state
*s
,
3946 int disk_idx
, int disks
)
3948 struct r5dev
*dev
= &sh
->dev
[disk_idx
];
3950 /* is the data in this block needed, and can we get it? */
3951 if (need_this_block(sh
, s
, disk_idx
, disks
)) {
3952 /* we would like to get this block, possibly by computing it,
3953 * otherwise read it if the backing disk is insync
3955 BUG_ON(test_bit(R5_Wantcompute
, &dev
->flags
));
3956 BUG_ON(test_bit(R5_Wantread
, &dev
->flags
));
3957 BUG_ON(sh
->batch_head
);
3960 * In the raid6 case if the only non-uptodate disk is P
3961 * then we already trusted P to compute the other failed
3962 * drives. It is safe to compute rather than re-read P.
3963 * In other cases we only compute blocks from failed
3964 * devices, otherwise check/repair might fail to detect
3965 * a real inconsistency.
3968 if ((s
->uptodate
== disks
- 1) &&
3969 ((sh
->qd_idx
>= 0 && sh
->pd_idx
== disk_idx
) ||
3970 (s
->failed
&& (disk_idx
== s
->failed_num
[0] ||
3971 disk_idx
== s
->failed_num
[1])))) {
3972 /* have disk failed, and we're requested to fetch it;
3975 pr_debug("Computing stripe %llu block %d\n",
3976 (unsigned long long)sh
->sector
, disk_idx
);
3977 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
3978 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
3979 set_bit(R5_Wantcompute
, &dev
->flags
);
3980 sh
->ops
.target
= disk_idx
;
3981 sh
->ops
.target2
= -1; /* no 2nd target */
3983 /* Careful: from this point on 'uptodate' is in the eye
3984 * of raid_run_ops which services 'compute' operations
3985 * before writes. R5_Wantcompute flags a block that will
3986 * be R5_UPTODATE by the time it is needed for a
3987 * subsequent operation.
3991 } else if (s
->uptodate
== disks
-2 && s
->failed
>= 2) {
3992 /* Computing 2-failure is *very* expensive; only
3993 * do it if failed >= 2
3996 for (other
= disks
; other
--; ) {
3997 if (other
== disk_idx
)
3999 if (!test_bit(R5_UPTODATE
,
4000 &sh
->dev
[other
].flags
))
4004 pr_debug("Computing stripe %llu blocks %d,%d\n",
4005 (unsigned long long)sh
->sector
,
4007 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
4008 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
4009 set_bit(R5_Wantcompute
, &sh
->dev
[disk_idx
].flags
);
4010 set_bit(R5_Wantcompute
, &sh
->dev
[other
].flags
);
4011 sh
->ops
.target
= disk_idx
;
4012 sh
->ops
.target2
= other
;
4016 } else if (test_bit(R5_Insync
, &dev
->flags
)) {
4017 set_bit(R5_LOCKED
, &dev
->flags
);
4018 set_bit(R5_Wantread
, &dev
->flags
);
4020 pr_debug("Reading block %d (sync=%d)\n",
4021 disk_idx
, s
->syncing
);
4029 * handle_stripe_fill - read or compute data to satisfy pending requests.
4031 static void handle_stripe_fill(struct stripe_head
*sh
,
4032 struct stripe_head_state
*s
,
4037 /* look for blocks to read/compute, skip this if a compute
4038 * is already in flight, or if the stripe contents are in the
4039 * midst of changing due to a write
4041 if (!test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
) && !sh
->check_state
&&
4042 !sh
->reconstruct_state
) {
4045 * For degraded stripe with data in journal, do not handle
4046 * read requests yet, instead, flush the stripe to raid
4047 * disks first, this avoids handling complex rmw of write
4048 * back cache (prexor with orig_page, and then xor with
4049 * page) in the read path
4051 if (s
->to_read
&& s
->injournal
&& s
->failed
) {
4052 if (test_bit(STRIPE_R5C_CACHING
, &sh
->state
))
4053 r5c_make_stripe_write_out(sh
);
4057 for (i
= disks
; i
--; )
4058 if (fetch_block(sh
, s
, i
, disks
))
4062 set_bit(STRIPE_HANDLE
, &sh
->state
);
4065 static void break_stripe_batch_list(struct stripe_head
*head_sh
,
4066 unsigned long handle_flags
);
4067 /* handle_stripe_clean_event
4068 * any written block on an uptodate or failed drive can be returned.
4069 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
4070 * never LOCKED, so we don't need to test 'failed' directly.
4072 static void handle_stripe_clean_event(struct r5conf
*conf
,
4073 struct stripe_head
*sh
, int disks
)
4077 int discard_pending
= 0;
4078 struct stripe_head
*head_sh
= sh
;
4079 bool do_endio
= false;
4081 for (i
= disks
; i
--; )
4082 if (sh
->dev
[i
].written
) {
4084 if (!test_bit(R5_LOCKED
, &dev
->flags
) &&
4085 (test_bit(R5_UPTODATE
, &dev
->flags
) ||
4086 test_bit(R5_Discard
, &dev
->flags
) ||
4087 test_bit(R5_SkipCopy
, &dev
->flags
))) {
4088 /* We can return any write requests */
4089 struct bio
*wbi
, *wbi2
;
4090 pr_debug("Return write for disc %d\n", i
);
4091 if (test_and_clear_bit(R5_Discard
, &dev
->flags
))
4092 clear_bit(R5_UPTODATE
, &dev
->flags
);
4093 if (test_and_clear_bit(R5_SkipCopy
, &dev
->flags
)) {
4094 WARN_ON(test_bit(R5_UPTODATE
, &dev
->flags
));
4099 dev
->page
= dev
->orig_page
;
4101 dev
->written
= NULL
;
4102 while (wbi
&& wbi
->bi_iter
.bi_sector
<
4103 dev
->sector
+ RAID5_STRIPE_SECTORS(conf
)) {
4104 wbi2
= r5_next_bio(conf
, wbi
, dev
->sector
);
4105 md_write_end(conf
->mddev
);
4109 md_bitmap_endwrite(conf
->mddev
->bitmap
, sh
->sector
,
4110 RAID5_STRIPE_SECTORS(conf
),
4111 !test_bit(STRIPE_DEGRADED
, &sh
->state
),
4113 if (head_sh
->batch_head
) {
4114 sh
= list_first_entry(&sh
->batch_list
,
4117 if (sh
!= head_sh
) {
4124 } else if (test_bit(R5_Discard
, &dev
->flags
))
4125 discard_pending
= 1;
4128 log_stripe_write_finished(sh
);
4130 if (!discard_pending
&&
4131 test_bit(R5_Discard
, &sh
->dev
[sh
->pd_idx
].flags
)) {
4133 clear_bit(R5_Discard
, &sh
->dev
[sh
->pd_idx
].flags
);
4134 clear_bit(R5_UPTODATE
, &sh
->dev
[sh
->pd_idx
].flags
);
4135 if (sh
->qd_idx
>= 0) {
4136 clear_bit(R5_Discard
, &sh
->dev
[sh
->qd_idx
].flags
);
4137 clear_bit(R5_UPTODATE
, &sh
->dev
[sh
->qd_idx
].flags
);
4139 /* now that discard is done we can proceed with any sync */
4140 clear_bit(STRIPE_DISCARD
, &sh
->state
);
4142 * SCSI discard will change some bio fields and the stripe has
4143 * no updated data, so remove it from hash list and the stripe
4144 * will be reinitialized
4147 hash
= sh
->hash_lock_index
;
4148 spin_lock_irq(conf
->hash_locks
+ hash
);
4150 spin_unlock_irq(conf
->hash_locks
+ hash
);
4151 if (head_sh
->batch_head
) {
4152 sh
= list_first_entry(&sh
->batch_list
,
4153 struct stripe_head
, batch_list
);
4159 if (test_bit(STRIPE_SYNC_REQUESTED
, &sh
->state
))
4160 set_bit(STRIPE_HANDLE
, &sh
->state
);
4164 if (test_and_clear_bit(STRIPE_FULL_WRITE
, &sh
->state
))
4165 if (atomic_dec_and_test(&conf
->pending_full_writes
))
4166 md_wakeup_thread(conf
->mddev
->thread
);
4168 if (head_sh
->batch_head
&& do_endio
)
4169 break_stripe_batch_list(head_sh
, STRIPE_EXPAND_SYNC_FLAGS
);
4173 * For RMW in write back cache, we need extra page in prexor to store the
4174 * old data. This page is stored in dev->orig_page.
4176 * This function checks whether we have data for prexor. The exact logic
4178 * R5_UPTODATE && (!R5_InJournal || R5_OrigPageUPTDODATE)
4180 static inline bool uptodate_for_rmw(struct r5dev
*dev
)
4182 return (test_bit(R5_UPTODATE
, &dev
->flags
)) &&
4183 (!test_bit(R5_InJournal
, &dev
->flags
) ||
4184 test_bit(R5_OrigPageUPTDODATE
, &dev
->flags
));
4187 static int handle_stripe_dirtying(struct r5conf
*conf
,
4188 struct stripe_head
*sh
,
4189 struct stripe_head_state
*s
,
4192 int rmw
= 0, rcw
= 0, i
;
4193 sector_t recovery_cp
= conf
->mddev
->recovery_cp
;
4195 /* Check whether resync is now happening or should start.
4196 * If yes, then the array is dirty (after unclean shutdown or
4197 * initial creation), so parity in some stripes might be inconsistent.
4198 * In this case, we need to always do reconstruct-write, to ensure
4199 * that in case of drive failure or read-error correction, we
4200 * generate correct data from the parity.
4202 if (conf
->rmw_level
== PARITY_DISABLE_RMW
||
4203 (recovery_cp
< MaxSector
&& sh
->sector
>= recovery_cp
&&
4205 /* Calculate the real rcw later - for now make it
4206 * look like rcw is cheaper
4209 pr_debug("force RCW rmw_level=%u, recovery_cp=%llu sh->sector=%llu\n",
4210 conf
->rmw_level
, (unsigned long long)recovery_cp
,
4211 (unsigned long long)sh
->sector
);
4212 } else for (i
= disks
; i
--; ) {
4213 /* would I have to read this buffer for read_modify_write */
4214 struct r5dev
*dev
= &sh
->dev
[i
];
4215 if (((dev
->towrite
&& !delay_towrite(conf
, dev
, s
)) ||
4216 i
== sh
->pd_idx
|| i
== sh
->qd_idx
||
4217 test_bit(R5_InJournal
, &dev
->flags
)) &&
4218 !test_bit(R5_LOCKED
, &dev
->flags
) &&
4219 !(uptodate_for_rmw(dev
) ||
4220 test_bit(R5_Wantcompute
, &dev
->flags
))) {
4221 if (test_bit(R5_Insync
, &dev
->flags
))
4224 rmw
+= 2*disks
; /* cannot read it */
4226 /* Would I have to read this buffer for reconstruct_write */
4227 if (!test_bit(R5_OVERWRITE
, &dev
->flags
) &&
4228 i
!= sh
->pd_idx
&& i
!= sh
->qd_idx
&&
4229 !test_bit(R5_LOCKED
, &dev
->flags
) &&
4230 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
4231 test_bit(R5_Wantcompute
, &dev
->flags
))) {
4232 if (test_bit(R5_Insync
, &dev
->flags
))
4239 pr_debug("for sector %llu state 0x%lx, rmw=%d rcw=%d\n",
4240 (unsigned long long)sh
->sector
, sh
->state
, rmw
, rcw
);
4241 set_bit(STRIPE_HANDLE
, &sh
->state
);
4242 if ((rmw
< rcw
|| (rmw
== rcw
&& conf
->rmw_level
== PARITY_PREFER_RMW
)) && rmw
> 0) {
4243 /* prefer read-modify-write, but need to get some data */
4244 if (conf
->mddev
->queue
)
4245 blk_add_trace_msg(conf
->mddev
->queue
,
4246 "raid5 rmw %llu %d",
4247 (unsigned long long)sh
->sector
, rmw
);
4248 for (i
= disks
; i
--; ) {
4249 struct r5dev
*dev
= &sh
->dev
[i
];
4250 if (test_bit(R5_InJournal
, &dev
->flags
) &&
4251 dev
->page
== dev
->orig_page
&&
4252 !test_bit(R5_LOCKED
, &sh
->dev
[sh
->pd_idx
].flags
)) {
4253 /* alloc page for prexor */
4254 struct page
*p
= alloc_page(GFP_NOIO
);
4262 * alloc_page() failed, try use
4263 * disk_info->extra_page
4265 if (!test_and_set_bit(R5C_EXTRA_PAGE_IN_USE
,
4266 &conf
->cache_state
)) {
4267 r5c_use_extra_page(sh
);
4271 /* extra_page in use, add to delayed_list */
4272 set_bit(STRIPE_DELAYED
, &sh
->state
);
4273 s
->waiting_extra_page
= 1;
4278 for (i
= disks
; i
--; ) {
4279 struct r5dev
*dev
= &sh
->dev
[i
];
4280 if (((dev
->towrite
&& !delay_towrite(conf
, dev
, s
)) ||
4281 i
== sh
->pd_idx
|| i
== sh
->qd_idx
||
4282 test_bit(R5_InJournal
, &dev
->flags
)) &&
4283 !test_bit(R5_LOCKED
, &dev
->flags
) &&
4284 !(uptodate_for_rmw(dev
) ||
4285 test_bit(R5_Wantcompute
, &dev
->flags
)) &&
4286 test_bit(R5_Insync
, &dev
->flags
)) {
4287 if (test_bit(STRIPE_PREREAD_ACTIVE
,
4289 pr_debug("Read_old block %d for r-m-w\n",
4291 set_bit(R5_LOCKED
, &dev
->flags
);
4292 set_bit(R5_Wantread
, &dev
->flags
);
4295 set_bit(STRIPE_DELAYED
, &sh
->state
);
4299 if ((rcw
< rmw
|| (rcw
== rmw
&& conf
->rmw_level
!= PARITY_PREFER_RMW
)) && rcw
> 0) {
4300 /* want reconstruct write, but need to get some data */
4303 for (i
= disks
; i
--; ) {
4304 struct r5dev
*dev
= &sh
->dev
[i
];
4305 if (!test_bit(R5_OVERWRITE
, &dev
->flags
) &&
4306 i
!= sh
->pd_idx
&& i
!= sh
->qd_idx
&&
4307 !test_bit(R5_LOCKED
, &dev
->flags
) &&
4308 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
4309 test_bit(R5_Wantcompute
, &dev
->flags
))) {
4311 if (test_bit(R5_Insync
, &dev
->flags
) &&
4312 test_bit(STRIPE_PREREAD_ACTIVE
,
4314 pr_debug("Read_old block "
4315 "%d for Reconstruct\n", i
);
4316 set_bit(R5_LOCKED
, &dev
->flags
);
4317 set_bit(R5_Wantread
, &dev
->flags
);
4321 set_bit(STRIPE_DELAYED
, &sh
->state
);
4324 if (rcw
&& conf
->mddev
->queue
)
4325 blk_add_trace_msg(conf
->mddev
->queue
, "raid5 rcw %llu %d %d %d",
4326 (unsigned long long)sh
->sector
,
4327 rcw
, qread
, test_bit(STRIPE_DELAYED
, &sh
->state
));
4330 if (rcw
> disks
&& rmw
> disks
&&
4331 !test_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
4332 set_bit(STRIPE_DELAYED
, &sh
->state
);
4334 /* now if nothing is locked, and if we have enough data,
4335 * we can start a write request
4337 /* since handle_stripe can be called at any time we need to handle the
4338 * case where a compute block operation has been submitted and then a
4339 * subsequent call wants to start a write request. raid_run_ops only
4340 * handles the case where compute block and reconstruct are requested
4341 * simultaneously. If this is not the case then new writes need to be
4342 * held off until the compute completes.
4344 if ((s
->req_compute
|| !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
)) &&
4345 (s
->locked
== 0 && (rcw
== 0 || rmw
== 0) &&
4346 !test_bit(STRIPE_BIT_DELAY
, &sh
->state
)))
4347 schedule_reconstruction(sh
, s
, rcw
== 0, 0);
4351 static void handle_parity_checks5(struct r5conf
*conf
, struct stripe_head
*sh
,
4352 struct stripe_head_state
*s
, int disks
)
4354 struct r5dev
*dev
= NULL
;
4356 BUG_ON(sh
->batch_head
);
4357 set_bit(STRIPE_HANDLE
, &sh
->state
);
4359 switch (sh
->check_state
) {
4360 case check_state_idle
:
4361 /* start a new check operation if there are no failures */
4362 if (s
->failed
== 0) {
4363 BUG_ON(s
->uptodate
!= disks
);
4364 sh
->check_state
= check_state_run
;
4365 set_bit(STRIPE_OP_CHECK
, &s
->ops_request
);
4366 clear_bit(R5_UPTODATE
, &sh
->dev
[sh
->pd_idx
].flags
);
4370 dev
= &sh
->dev
[s
->failed_num
[0]];
4372 case check_state_compute_result
:
4373 sh
->check_state
= check_state_idle
;
4375 dev
= &sh
->dev
[sh
->pd_idx
];
4377 /* check that a write has not made the stripe insync */
4378 if (test_bit(STRIPE_INSYNC
, &sh
->state
))
4381 /* either failed parity check, or recovery is happening */
4382 BUG_ON(!test_bit(R5_UPTODATE
, &dev
->flags
));
4383 BUG_ON(s
->uptodate
!= disks
);
4385 set_bit(R5_LOCKED
, &dev
->flags
);
4387 set_bit(R5_Wantwrite
, &dev
->flags
);
4389 clear_bit(STRIPE_DEGRADED
, &sh
->state
);
4390 set_bit(STRIPE_INSYNC
, &sh
->state
);
4392 case check_state_run
:
4393 break; /* we will be called again upon completion */
4394 case check_state_check_result
:
4395 sh
->check_state
= check_state_idle
;
4397 /* if a failure occurred during the check operation, leave
4398 * STRIPE_INSYNC not set and let the stripe be handled again
4403 /* handle a successful check operation, if parity is correct
4404 * we are done. Otherwise update the mismatch count and repair
4405 * parity if !MD_RECOVERY_CHECK
4407 if ((sh
->ops
.zero_sum_result
& SUM_CHECK_P_RESULT
) == 0)
4408 /* parity is correct (on disc,
4409 * not in buffer any more)
4411 set_bit(STRIPE_INSYNC
, &sh
->state
);
4413 atomic64_add(RAID5_STRIPE_SECTORS(conf
), &conf
->mddev
->resync_mismatches
);
4414 if (test_bit(MD_RECOVERY_CHECK
, &conf
->mddev
->recovery
)) {
4415 /* don't try to repair!! */
4416 set_bit(STRIPE_INSYNC
, &sh
->state
);
4417 pr_warn_ratelimited("%s: mismatch sector in range "
4418 "%llu-%llu\n", mdname(conf
->mddev
),
4419 (unsigned long long) sh
->sector
,
4420 (unsigned long long) sh
->sector
+
4421 RAID5_STRIPE_SECTORS(conf
));
4423 sh
->check_state
= check_state_compute_run
;
4424 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
4425 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
4426 set_bit(R5_Wantcompute
,
4427 &sh
->dev
[sh
->pd_idx
].flags
);
4428 sh
->ops
.target
= sh
->pd_idx
;
4429 sh
->ops
.target2
= -1;
4434 case check_state_compute_run
:
4437 pr_err("%s: unknown check_state: %d sector: %llu\n",
4438 __func__
, sh
->check_state
,
4439 (unsigned long long) sh
->sector
);
4444 static void handle_parity_checks6(struct r5conf
*conf
, struct stripe_head
*sh
,
4445 struct stripe_head_state
*s
,
4448 int pd_idx
= sh
->pd_idx
;
4449 int qd_idx
= sh
->qd_idx
;
4452 BUG_ON(sh
->batch_head
);
4453 set_bit(STRIPE_HANDLE
, &sh
->state
);
4455 BUG_ON(s
->failed
> 2);
4457 /* Want to check and possibly repair P and Q.
4458 * However there could be one 'failed' device, in which
4459 * case we can only check one of them, possibly using the
4460 * other to generate missing data
4463 switch (sh
->check_state
) {
4464 case check_state_idle
:
4465 /* start a new check operation if there are < 2 failures */
4466 if (s
->failed
== s
->q_failed
) {
4467 /* The only possible failed device holds Q, so it
4468 * makes sense to check P (If anything else were failed,
4469 * we would have used P to recreate it).
4471 sh
->check_state
= check_state_run
;
4473 if (!s
->q_failed
&& s
->failed
< 2) {
4474 /* Q is not failed, and we didn't use it to generate
4475 * anything, so it makes sense to check it
4477 if (sh
->check_state
== check_state_run
)
4478 sh
->check_state
= check_state_run_pq
;
4480 sh
->check_state
= check_state_run_q
;
4483 /* discard potentially stale zero_sum_result */
4484 sh
->ops
.zero_sum_result
= 0;
4486 if (sh
->check_state
== check_state_run
) {
4487 /* async_xor_zero_sum destroys the contents of P */
4488 clear_bit(R5_UPTODATE
, &sh
->dev
[pd_idx
].flags
);
4491 if (sh
->check_state
>= check_state_run
&&
4492 sh
->check_state
<= check_state_run_pq
) {
4493 /* async_syndrome_zero_sum preserves P and Q, so
4494 * no need to mark them !uptodate here
4496 set_bit(STRIPE_OP_CHECK
, &s
->ops_request
);
4500 /* we have 2-disk failure */
4501 BUG_ON(s
->failed
!= 2);
4503 case check_state_compute_result
:
4504 sh
->check_state
= check_state_idle
;
4506 /* check that a write has not made the stripe insync */
4507 if (test_bit(STRIPE_INSYNC
, &sh
->state
))
4510 /* now write out any block on a failed drive,
4511 * or P or Q if they were recomputed
4514 if (s
->failed
== 2) {
4515 dev
= &sh
->dev
[s
->failed_num
[1]];
4517 set_bit(R5_LOCKED
, &dev
->flags
);
4518 set_bit(R5_Wantwrite
, &dev
->flags
);
4520 if (s
->failed
>= 1) {
4521 dev
= &sh
->dev
[s
->failed_num
[0]];
4523 set_bit(R5_LOCKED
, &dev
->flags
);
4524 set_bit(R5_Wantwrite
, &dev
->flags
);
4526 if (sh
->ops
.zero_sum_result
& SUM_CHECK_P_RESULT
) {
4527 dev
= &sh
->dev
[pd_idx
];
4529 set_bit(R5_LOCKED
, &dev
->flags
);
4530 set_bit(R5_Wantwrite
, &dev
->flags
);
4532 if (sh
->ops
.zero_sum_result
& SUM_CHECK_Q_RESULT
) {
4533 dev
= &sh
->dev
[qd_idx
];
4535 set_bit(R5_LOCKED
, &dev
->flags
);
4536 set_bit(R5_Wantwrite
, &dev
->flags
);
4538 if (WARN_ONCE(dev
&& !test_bit(R5_UPTODATE
, &dev
->flags
),
4539 "%s: disk%td not up to date\n",
4540 mdname(conf
->mddev
),
4541 dev
- (struct r5dev
*) &sh
->dev
)) {
4542 clear_bit(R5_LOCKED
, &dev
->flags
);
4543 clear_bit(R5_Wantwrite
, &dev
->flags
);
4546 clear_bit(STRIPE_DEGRADED
, &sh
->state
);
4548 set_bit(STRIPE_INSYNC
, &sh
->state
);
4550 case check_state_run
:
4551 case check_state_run_q
:
4552 case check_state_run_pq
:
4553 break; /* we will be called again upon completion */
4554 case check_state_check_result
:
4555 sh
->check_state
= check_state_idle
;
4557 /* handle a successful check operation, if parity is correct
4558 * we are done. Otherwise update the mismatch count and repair
4559 * parity if !MD_RECOVERY_CHECK
4561 if (sh
->ops
.zero_sum_result
== 0) {
4562 /* both parities are correct */
4564 set_bit(STRIPE_INSYNC
, &sh
->state
);
4566 /* in contrast to the raid5 case we can validate
4567 * parity, but still have a failure to write
4570 sh
->check_state
= check_state_compute_result
;
4571 /* Returning at this point means that we may go
4572 * off and bring p and/or q uptodate again so
4573 * we make sure to check zero_sum_result again
4574 * to verify if p or q need writeback
4578 atomic64_add(RAID5_STRIPE_SECTORS(conf
), &conf
->mddev
->resync_mismatches
);
4579 if (test_bit(MD_RECOVERY_CHECK
, &conf
->mddev
->recovery
)) {
4580 /* don't try to repair!! */
4581 set_bit(STRIPE_INSYNC
, &sh
->state
);
4582 pr_warn_ratelimited("%s: mismatch sector in range "
4583 "%llu-%llu\n", mdname(conf
->mddev
),
4584 (unsigned long long) sh
->sector
,
4585 (unsigned long long) sh
->sector
+
4586 RAID5_STRIPE_SECTORS(conf
));
4588 int *target
= &sh
->ops
.target
;
4590 sh
->ops
.target
= -1;
4591 sh
->ops
.target2
= -1;
4592 sh
->check_state
= check_state_compute_run
;
4593 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
4594 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
4595 if (sh
->ops
.zero_sum_result
& SUM_CHECK_P_RESULT
) {
4596 set_bit(R5_Wantcompute
,
4597 &sh
->dev
[pd_idx
].flags
);
4599 target
= &sh
->ops
.target2
;
4602 if (sh
->ops
.zero_sum_result
& SUM_CHECK_Q_RESULT
) {
4603 set_bit(R5_Wantcompute
,
4604 &sh
->dev
[qd_idx
].flags
);
4611 case check_state_compute_run
:
4614 pr_warn("%s: unknown check_state: %d sector: %llu\n",
4615 __func__
, sh
->check_state
,
4616 (unsigned long long) sh
->sector
);
4621 static void handle_stripe_expansion(struct r5conf
*conf
, struct stripe_head
*sh
)
4625 /* We have read all the blocks in this stripe and now we need to
4626 * copy some of them into a target stripe for expand.
4628 struct dma_async_tx_descriptor
*tx
= NULL
;
4629 BUG_ON(sh
->batch_head
);
4630 clear_bit(STRIPE_EXPAND_SOURCE
, &sh
->state
);
4631 for (i
= 0; i
< sh
->disks
; i
++)
4632 if (i
!= sh
->pd_idx
&& i
!= sh
->qd_idx
) {
4634 struct stripe_head
*sh2
;
4635 struct async_submit_ctl submit
;
4637 sector_t bn
= raid5_compute_blocknr(sh
, i
, 1);
4638 sector_t s
= raid5_compute_sector(conf
, bn
, 0,
4640 sh2
= raid5_get_active_stripe(conf
, NULL
, s
,
4641 R5_GAS_NOBLOCK
| R5_GAS_NOQUIESCE
);
4643 /* so far only the early blocks of this stripe
4644 * have been requested. When later blocks
4645 * get requested, we will try again
4648 if (!test_bit(STRIPE_EXPANDING
, &sh2
->state
) ||
4649 test_bit(R5_Expanded
, &sh2
->dev
[dd_idx
].flags
)) {
4650 /* must have already done this block */
4651 raid5_release_stripe(sh2
);
4655 /* place all the copies on one channel */
4656 init_async_submit(&submit
, 0, tx
, NULL
, NULL
, NULL
);
4657 tx
= async_memcpy(sh2
->dev
[dd_idx
].page
,
4658 sh
->dev
[i
].page
, sh2
->dev
[dd_idx
].offset
,
4659 sh
->dev
[i
].offset
, RAID5_STRIPE_SIZE(conf
),
4662 set_bit(R5_Expanded
, &sh2
->dev
[dd_idx
].flags
);
4663 set_bit(R5_UPTODATE
, &sh2
->dev
[dd_idx
].flags
);
4664 for (j
= 0; j
< conf
->raid_disks
; j
++)
4665 if (j
!= sh2
->pd_idx
&&
4667 !test_bit(R5_Expanded
, &sh2
->dev
[j
].flags
))
4669 if (j
== conf
->raid_disks
) {
4670 set_bit(STRIPE_EXPAND_READY
, &sh2
->state
);
4671 set_bit(STRIPE_HANDLE
, &sh2
->state
);
4673 raid5_release_stripe(sh2
);
4676 /* done submitting copies, wait for them to complete */
4677 async_tx_quiesce(&tx
);
4681 * handle_stripe - do things to a stripe.
4683 * We lock the stripe by setting STRIPE_ACTIVE and then examine the
4684 * state of various bits to see what needs to be done.
4686 * return some read requests which now have data
4687 * return some write requests which are safely on storage
4688 * schedule a read on some buffers
4689 * schedule a write of some buffers
4690 * return confirmation of parity correctness
4694 static void analyse_stripe(struct stripe_head
*sh
, struct stripe_head_state
*s
)
4696 struct r5conf
*conf
= sh
->raid_conf
;
4697 int disks
= sh
->disks
;
4700 int do_recovery
= 0;
4702 memset(s
, 0, sizeof(*s
));
4704 s
->expanding
= test_bit(STRIPE_EXPAND_SOURCE
, &sh
->state
) && !sh
->batch_head
;
4705 s
->expanded
= test_bit(STRIPE_EXPAND_READY
, &sh
->state
) && !sh
->batch_head
;
4706 s
->failed_num
[0] = -1;
4707 s
->failed_num
[1] = -1;
4708 s
->log_failed
= r5l_log_disk_error(conf
);
4710 /* Now to look around and see what can be done */
4712 for (i
=disks
; i
--; ) {
4713 struct md_rdev
*rdev
;
4720 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
4722 dev
->toread
, dev
->towrite
, dev
->written
);
4723 /* maybe we can reply to a read
4725 * new wantfill requests are only permitted while
4726 * ops_complete_biofill is guaranteed to be inactive
4728 if (test_bit(R5_UPTODATE
, &dev
->flags
) && dev
->toread
&&
4729 !test_bit(STRIPE_BIOFILL_RUN
, &sh
->state
))
4730 set_bit(R5_Wantfill
, &dev
->flags
);
4732 /* now count some things */
4733 if (test_bit(R5_LOCKED
, &dev
->flags
))
4735 if (test_bit(R5_UPTODATE
, &dev
->flags
))
4737 if (test_bit(R5_Wantcompute
, &dev
->flags
)) {
4739 BUG_ON(s
->compute
> 2);
4742 if (test_bit(R5_Wantfill
, &dev
->flags
))
4744 else if (dev
->toread
)
4748 if (!test_bit(R5_OVERWRITE
, &dev
->flags
))
4753 /* Prefer to use the replacement for reads, but only
4754 * if it is recovered enough and has no bad blocks.
4756 rdev
= rcu_dereference(conf
->disks
[i
].replacement
);
4757 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
) &&
4758 rdev
->recovery_offset
>= sh
->sector
+ RAID5_STRIPE_SECTORS(conf
) &&
4759 !is_badblock(rdev
, sh
->sector
, RAID5_STRIPE_SECTORS(conf
),
4760 &first_bad
, &bad_sectors
))
4761 set_bit(R5_ReadRepl
, &dev
->flags
);
4763 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
))
4764 set_bit(R5_NeedReplace
, &dev
->flags
);
4766 clear_bit(R5_NeedReplace
, &dev
->flags
);
4767 rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
4768 clear_bit(R5_ReadRepl
, &dev
->flags
);
4770 if (rdev
&& test_bit(Faulty
, &rdev
->flags
))
4773 is_bad
= is_badblock(rdev
, sh
->sector
, RAID5_STRIPE_SECTORS(conf
),
4774 &first_bad
, &bad_sectors
);
4775 if (s
->blocked_rdev
== NULL
4776 && (test_bit(Blocked
, &rdev
->flags
)
4779 set_bit(BlockedBadBlocks
,
4781 s
->blocked_rdev
= rdev
;
4782 atomic_inc(&rdev
->nr_pending
);
4785 clear_bit(R5_Insync
, &dev
->flags
);
4789 /* also not in-sync */
4790 if (!test_bit(WriteErrorSeen
, &rdev
->flags
) &&
4791 test_bit(R5_UPTODATE
, &dev
->flags
)) {
4792 /* treat as in-sync, but with a read error
4793 * which we can now try to correct
4795 set_bit(R5_Insync
, &dev
->flags
);
4796 set_bit(R5_ReadError
, &dev
->flags
);
4798 } else if (test_bit(In_sync
, &rdev
->flags
))
4799 set_bit(R5_Insync
, &dev
->flags
);
4800 else if (sh
->sector
+ RAID5_STRIPE_SECTORS(conf
) <= rdev
->recovery_offset
)
4801 /* in sync if before recovery_offset */
4802 set_bit(R5_Insync
, &dev
->flags
);
4803 else if (test_bit(R5_UPTODATE
, &dev
->flags
) &&
4804 test_bit(R5_Expanded
, &dev
->flags
))
4805 /* If we've reshaped into here, we assume it is Insync.
4806 * We will shortly update recovery_offset to make
4809 set_bit(R5_Insync
, &dev
->flags
);
4811 if (test_bit(R5_WriteError
, &dev
->flags
)) {
4812 /* This flag does not apply to '.replacement'
4813 * only to .rdev, so make sure to check that*/
4814 struct md_rdev
*rdev2
= rcu_dereference(
4815 conf
->disks
[i
].rdev
);
4817 clear_bit(R5_Insync
, &dev
->flags
);
4818 if (rdev2
&& !test_bit(Faulty
, &rdev2
->flags
)) {
4819 s
->handle_bad_blocks
= 1;
4820 atomic_inc(&rdev2
->nr_pending
);
4822 clear_bit(R5_WriteError
, &dev
->flags
);
4824 if (test_bit(R5_MadeGood
, &dev
->flags
)) {
4825 /* This flag does not apply to '.replacement'
4826 * only to .rdev, so make sure to check that*/
4827 struct md_rdev
*rdev2
= rcu_dereference(
4828 conf
->disks
[i
].rdev
);
4829 if (rdev2
&& !test_bit(Faulty
, &rdev2
->flags
)) {
4830 s
->handle_bad_blocks
= 1;
4831 atomic_inc(&rdev2
->nr_pending
);
4833 clear_bit(R5_MadeGood
, &dev
->flags
);
4835 if (test_bit(R5_MadeGoodRepl
, &dev
->flags
)) {
4836 struct md_rdev
*rdev2
= rcu_dereference(
4837 conf
->disks
[i
].replacement
);
4838 if (rdev2
&& !test_bit(Faulty
, &rdev2
->flags
)) {
4839 s
->handle_bad_blocks
= 1;
4840 atomic_inc(&rdev2
->nr_pending
);
4842 clear_bit(R5_MadeGoodRepl
, &dev
->flags
);
4844 if (!test_bit(R5_Insync
, &dev
->flags
)) {
4845 /* The ReadError flag will just be confusing now */
4846 clear_bit(R5_ReadError
, &dev
->flags
);
4847 clear_bit(R5_ReWrite
, &dev
->flags
);
4849 if (test_bit(R5_ReadError
, &dev
->flags
))
4850 clear_bit(R5_Insync
, &dev
->flags
);
4851 if (!test_bit(R5_Insync
, &dev
->flags
)) {
4853 s
->failed_num
[s
->failed
] = i
;
4855 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
))
4858 rdev
= rcu_dereference(
4859 conf
->disks
[i
].replacement
);
4860 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
))
4865 if (test_bit(R5_InJournal
, &dev
->flags
))
4867 if (test_bit(R5_InJournal
, &dev
->flags
) && dev
->written
)
4870 if (test_bit(STRIPE_SYNCING
, &sh
->state
)) {
4871 /* If there is a failed device being replaced,
4872 * we must be recovering.
4873 * else if we are after recovery_cp, we must be syncing
4874 * else if MD_RECOVERY_REQUESTED is set, we also are syncing.
4875 * else we can only be replacing
4876 * sync and recovery both need to read all devices, and so
4877 * use the same flag.
4880 sh
->sector
>= conf
->mddev
->recovery_cp
||
4881 test_bit(MD_RECOVERY_REQUESTED
, &(conf
->mddev
->recovery
)))
4890 * Return '1' if this is a member of batch, or '0' if it is a lone stripe or
4891 * a head which can now be handled.
4893 static int clear_batch_ready(struct stripe_head
*sh
)
4895 struct stripe_head
*tmp
;
4896 if (!test_and_clear_bit(STRIPE_BATCH_READY
, &sh
->state
))
4897 return (sh
->batch_head
&& sh
->batch_head
!= sh
);
4898 spin_lock(&sh
->stripe_lock
);
4899 if (!sh
->batch_head
) {
4900 spin_unlock(&sh
->stripe_lock
);
4905 * this stripe could be added to a batch list before we check
4906 * BATCH_READY, skips it
4908 if (sh
->batch_head
!= sh
) {
4909 spin_unlock(&sh
->stripe_lock
);
4912 spin_lock(&sh
->batch_lock
);
4913 list_for_each_entry(tmp
, &sh
->batch_list
, batch_list
)
4914 clear_bit(STRIPE_BATCH_READY
, &tmp
->state
);
4915 spin_unlock(&sh
->batch_lock
);
4916 spin_unlock(&sh
->stripe_lock
);
4919 * BATCH_READY is cleared, no new stripes can be added.
4920 * batch_list can be accessed without lock
4925 static void break_stripe_batch_list(struct stripe_head
*head_sh
,
4926 unsigned long handle_flags
)
4928 struct stripe_head
*sh
, *next
;
4932 list_for_each_entry_safe(sh
, next
, &head_sh
->batch_list
, batch_list
) {
4934 list_del_init(&sh
->batch_list
);
4936 WARN_ONCE(sh
->state
& ((1 << STRIPE_ACTIVE
) |
4937 (1 << STRIPE_SYNCING
) |
4938 (1 << STRIPE_REPLACED
) |
4939 (1 << STRIPE_DELAYED
) |
4940 (1 << STRIPE_BIT_DELAY
) |
4941 (1 << STRIPE_FULL_WRITE
) |
4942 (1 << STRIPE_BIOFILL_RUN
) |
4943 (1 << STRIPE_COMPUTE_RUN
) |
4944 (1 << STRIPE_DISCARD
) |
4945 (1 << STRIPE_BATCH_READY
) |
4946 (1 << STRIPE_BATCH_ERR
) |
4947 (1 << STRIPE_BITMAP_PENDING
)),
4948 "stripe state: %lx\n", sh
->state
);
4949 WARN_ONCE(head_sh
->state
& ((1 << STRIPE_DISCARD
) |
4950 (1 << STRIPE_REPLACED
)),
4951 "head stripe state: %lx\n", head_sh
->state
);
4953 set_mask_bits(&sh
->state
, ~(STRIPE_EXPAND_SYNC_FLAGS
|
4954 (1 << STRIPE_PREREAD_ACTIVE
) |
4955 (1 << STRIPE_DEGRADED
) |
4956 (1 << STRIPE_ON_UNPLUG_LIST
)),
4957 head_sh
->state
& (1 << STRIPE_INSYNC
));
4959 sh
->check_state
= head_sh
->check_state
;
4960 sh
->reconstruct_state
= head_sh
->reconstruct_state
;
4961 spin_lock_irq(&sh
->stripe_lock
);
4962 sh
->batch_head
= NULL
;
4963 spin_unlock_irq(&sh
->stripe_lock
);
4964 for (i
= 0; i
< sh
->disks
; i
++) {
4965 if (test_and_clear_bit(R5_Overlap
, &sh
->dev
[i
].flags
))
4967 sh
->dev
[i
].flags
= head_sh
->dev
[i
].flags
&
4968 (~((1 << R5_WriteError
) | (1 << R5_Overlap
)));
4970 if (handle_flags
== 0 ||
4971 sh
->state
& handle_flags
)
4972 set_bit(STRIPE_HANDLE
, &sh
->state
);
4973 raid5_release_stripe(sh
);
4975 spin_lock_irq(&head_sh
->stripe_lock
);
4976 head_sh
->batch_head
= NULL
;
4977 spin_unlock_irq(&head_sh
->stripe_lock
);
4978 for (i
= 0; i
< head_sh
->disks
; i
++)
4979 if (test_and_clear_bit(R5_Overlap
, &head_sh
->dev
[i
].flags
))
4981 if (head_sh
->state
& handle_flags
)
4982 set_bit(STRIPE_HANDLE
, &head_sh
->state
);
4985 wake_up(&head_sh
->raid_conf
->wait_for_overlap
);
4988 static void handle_stripe(struct stripe_head
*sh
)
4990 struct stripe_head_state s
;
4991 struct r5conf
*conf
= sh
->raid_conf
;
4994 int disks
= sh
->disks
;
4995 struct r5dev
*pdev
, *qdev
;
4997 clear_bit(STRIPE_HANDLE
, &sh
->state
);
5000 * handle_stripe should not continue handle the batched stripe, only
5001 * the head of batch list or lone stripe can continue. Otherwise we
5002 * could see break_stripe_batch_list warns about the STRIPE_ACTIVE
5003 * is set for the batched stripe.
5005 if (clear_batch_ready(sh
))
5008 if (test_and_set_bit_lock(STRIPE_ACTIVE
, &sh
->state
)) {
5009 /* already being handled, ensure it gets handled
5010 * again when current action finishes */
5011 set_bit(STRIPE_HANDLE
, &sh
->state
);
5015 if (test_and_clear_bit(STRIPE_BATCH_ERR
, &sh
->state
))
5016 break_stripe_batch_list(sh
, 0);
5018 if (test_bit(STRIPE_SYNC_REQUESTED
, &sh
->state
) && !sh
->batch_head
) {
5019 spin_lock(&sh
->stripe_lock
);
5021 * Cannot process 'sync' concurrently with 'discard'.
5022 * Flush data in r5cache before 'sync'.
5024 if (!test_bit(STRIPE_R5C_PARTIAL_STRIPE
, &sh
->state
) &&
5025 !test_bit(STRIPE_R5C_FULL_STRIPE
, &sh
->state
) &&
5026 !test_bit(STRIPE_DISCARD
, &sh
->state
) &&
5027 test_and_clear_bit(STRIPE_SYNC_REQUESTED
, &sh
->state
)) {
5028 set_bit(STRIPE_SYNCING
, &sh
->state
);
5029 clear_bit(STRIPE_INSYNC
, &sh
->state
);
5030 clear_bit(STRIPE_REPLACED
, &sh
->state
);
5032 spin_unlock(&sh
->stripe_lock
);
5034 clear_bit(STRIPE_DELAYED
, &sh
->state
);
5036 pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
5037 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
5038 (unsigned long long)sh
->sector
, sh
->state
,
5039 atomic_read(&sh
->count
), sh
->pd_idx
, sh
->qd_idx
,
5040 sh
->check_state
, sh
->reconstruct_state
);
5042 analyse_stripe(sh
, &s
);
5044 if (test_bit(STRIPE_LOG_TRAPPED
, &sh
->state
))
5047 if (s
.handle_bad_blocks
||
5048 test_bit(MD_SB_CHANGE_PENDING
, &conf
->mddev
->sb_flags
)) {
5049 set_bit(STRIPE_HANDLE
, &sh
->state
);
5053 if (unlikely(s
.blocked_rdev
)) {
5054 if (s
.syncing
|| s
.expanding
|| s
.expanded
||
5055 s
.replacing
|| s
.to_write
|| s
.written
) {
5056 set_bit(STRIPE_HANDLE
, &sh
->state
);
5059 /* There is nothing for the blocked_rdev to block */
5060 rdev_dec_pending(s
.blocked_rdev
, conf
->mddev
);
5061 s
.blocked_rdev
= NULL
;
5064 if (s
.to_fill
&& !test_bit(STRIPE_BIOFILL_RUN
, &sh
->state
)) {
5065 set_bit(STRIPE_OP_BIOFILL
, &s
.ops_request
);
5066 set_bit(STRIPE_BIOFILL_RUN
, &sh
->state
);
5069 pr_debug("locked=%d uptodate=%d to_read=%d"
5070 " to_write=%d failed=%d failed_num=%d,%d\n",
5071 s
.locked
, s
.uptodate
, s
.to_read
, s
.to_write
, s
.failed
,
5072 s
.failed_num
[0], s
.failed_num
[1]);
5074 * check if the array has lost more than max_degraded devices and,
5075 * if so, some requests might need to be failed.
5077 * When journal device failed (log_failed), we will only process
5078 * the stripe if there is data need write to raid disks
5080 if (s
.failed
> conf
->max_degraded
||
5081 (s
.log_failed
&& s
.injournal
== 0)) {
5082 sh
->check_state
= 0;
5083 sh
->reconstruct_state
= 0;
5084 break_stripe_batch_list(sh
, 0);
5085 if (s
.to_read
+s
.to_write
+s
.written
)
5086 handle_failed_stripe(conf
, sh
, &s
, disks
);
5087 if (s
.syncing
+ s
.replacing
)
5088 handle_failed_sync(conf
, sh
, &s
);
5091 /* Now we check to see if any write operations have recently
5095 if (sh
->reconstruct_state
== reconstruct_state_prexor_drain_result
)
5097 if (sh
->reconstruct_state
== reconstruct_state_drain_result
||
5098 sh
->reconstruct_state
== reconstruct_state_prexor_drain_result
) {
5099 sh
->reconstruct_state
= reconstruct_state_idle
;
5101 /* All the 'written' buffers and the parity block are ready to
5102 * be written back to disk
5104 BUG_ON(!test_bit(R5_UPTODATE
, &sh
->dev
[sh
->pd_idx
].flags
) &&
5105 !test_bit(R5_Discard
, &sh
->dev
[sh
->pd_idx
].flags
));
5106 BUG_ON(sh
->qd_idx
>= 0 &&
5107 !test_bit(R5_UPTODATE
, &sh
->dev
[sh
->qd_idx
].flags
) &&
5108 !test_bit(R5_Discard
, &sh
->dev
[sh
->qd_idx
].flags
));
5109 for (i
= disks
; i
--; ) {
5110 struct r5dev
*dev
= &sh
->dev
[i
];
5111 if (test_bit(R5_LOCKED
, &dev
->flags
) &&
5112 (i
== sh
->pd_idx
|| i
== sh
->qd_idx
||
5113 dev
->written
|| test_bit(R5_InJournal
,
5115 pr_debug("Writing block %d\n", i
);
5116 set_bit(R5_Wantwrite
, &dev
->flags
);
5121 if (!test_bit(R5_Insync
, &dev
->flags
) ||
5122 ((i
== sh
->pd_idx
|| i
== sh
->qd_idx
) &&
5124 set_bit(STRIPE_INSYNC
, &sh
->state
);
5127 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
5128 s
.dec_preread_active
= 1;
5132 * might be able to return some write requests if the parity blocks
5133 * are safe, or on a failed drive
5135 pdev
= &sh
->dev
[sh
->pd_idx
];
5136 s
.p_failed
= (s
.failed
>= 1 && s
.failed_num
[0] == sh
->pd_idx
)
5137 || (s
.failed
>= 2 && s
.failed_num
[1] == sh
->pd_idx
);
5138 qdev
= &sh
->dev
[sh
->qd_idx
];
5139 s
.q_failed
= (s
.failed
>= 1 && s
.failed_num
[0] == sh
->qd_idx
)
5140 || (s
.failed
>= 2 && s
.failed_num
[1] == sh
->qd_idx
)
5144 (s
.p_failed
|| ((test_bit(R5_Insync
, &pdev
->flags
)
5145 && !test_bit(R5_LOCKED
, &pdev
->flags
)
5146 && (test_bit(R5_UPTODATE
, &pdev
->flags
) ||
5147 test_bit(R5_Discard
, &pdev
->flags
))))) &&
5148 (s
.q_failed
|| ((test_bit(R5_Insync
, &qdev
->flags
)
5149 && !test_bit(R5_LOCKED
, &qdev
->flags
)
5150 && (test_bit(R5_UPTODATE
, &qdev
->flags
) ||
5151 test_bit(R5_Discard
, &qdev
->flags
))))))
5152 handle_stripe_clean_event(conf
, sh
, disks
);
5155 r5c_handle_cached_data_endio(conf
, sh
, disks
);
5156 log_stripe_write_finished(sh
);
5158 /* Now we might consider reading some blocks, either to check/generate
5159 * parity, or to satisfy requests
5160 * or to load a block that is being partially written.
5162 if (s
.to_read
|| s
.non_overwrite
5163 || (s
.to_write
&& s
.failed
)
5164 || (s
.syncing
&& (s
.uptodate
+ s
.compute
< disks
))
5167 handle_stripe_fill(sh
, &s
, disks
);
5170 * When the stripe finishes full journal write cycle (write to journal
5171 * and raid disk), this is the clean up procedure so it is ready for
5174 r5c_finish_stripe_write_out(conf
, sh
, &s
);
5177 * Now to consider new write requests, cache write back and what else,
5178 * if anything should be read. We do not handle new writes when:
5179 * 1/ A 'write' operation (copy+xor) is already in flight.
5180 * 2/ A 'check' operation is in flight, as it may clobber the parity
5182 * 3/ A r5c cache log write is in flight.
5185 if (!sh
->reconstruct_state
&& !sh
->check_state
&& !sh
->log_io
) {
5186 if (!r5c_is_writeback(conf
->log
)) {
5188 handle_stripe_dirtying(conf
, sh
, &s
, disks
);
5189 } else { /* write back cache */
5192 /* First, try handle writes in caching phase */
5194 ret
= r5c_try_caching_write(conf
, sh
, &s
,
5197 * If caching phase failed: ret == -EAGAIN
5199 * stripe under reclaim: !caching && injournal
5201 * fall back to handle_stripe_dirtying()
5203 if (ret
== -EAGAIN
||
5204 /* stripe under reclaim: !caching && injournal */
5205 (!test_bit(STRIPE_R5C_CACHING
, &sh
->state
) &&
5207 ret
= handle_stripe_dirtying(conf
, sh
, &s
,
5215 /* maybe we need to check and possibly fix the parity for this stripe
5216 * Any reads will already have been scheduled, so we just see if enough
5217 * data is available. The parity check is held off while parity
5218 * dependent operations are in flight.
5220 if (sh
->check_state
||
5221 (s
.syncing
&& s
.locked
== 0 &&
5222 !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
) &&
5223 !test_bit(STRIPE_INSYNC
, &sh
->state
))) {
5224 if (conf
->level
== 6)
5225 handle_parity_checks6(conf
, sh
, &s
, disks
);
5227 handle_parity_checks5(conf
, sh
, &s
, disks
);
5230 if ((s
.replacing
|| s
.syncing
) && s
.locked
== 0
5231 && !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
)
5232 && !test_bit(STRIPE_REPLACED
, &sh
->state
)) {
5233 /* Write out to replacement devices where possible */
5234 for (i
= 0; i
< conf
->raid_disks
; i
++)
5235 if (test_bit(R5_NeedReplace
, &sh
->dev
[i
].flags
)) {
5236 WARN_ON(!test_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
));
5237 set_bit(R5_WantReplace
, &sh
->dev
[i
].flags
);
5238 set_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
5242 set_bit(STRIPE_INSYNC
, &sh
->state
);
5243 set_bit(STRIPE_REPLACED
, &sh
->state
);
5245 if ((s
.syncing
|| s
.replacing
) && s
.locked
== 0 &&
5246 !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
) &&
5247 test_bit(STRIPE_INSYNC
, &sh
->state
)) {
5248 md_done_sync(conf
->mddev
, RAID5_STRIPE_SECTORS(conf
), 1);
5249 clear_bit(STRIPE_SYNCING
, &sh
->state
);
5250 if (test_and_clear_bit(R5_Overlap
, &sh
->dev
[sh
->pd_idx
].flags
))
5251 wake_up(&conf
->wait_for_overlap
);
5254 /* If the failed drives are just a ReadError, then we might need
5255 * to progress the repair/check process
5257 if (s
.failed
<= conf
->max_degraded
&& !conf
->mddev
->ro
)
5258 for (i
= 0; i
< s
.failed
; i
++) {
5259 struct r5dev
*dev
= &sh
->dev
[s
.failed_num
[i
]];
5260 if (test_bit(R5_ReadError
, &dev
->flags
)
5261 && !test_bit(R5_LOCKED
, &dev
->flags
)
5262 && test_bit(R5_UPTODATE
, &dev
->flags
)
5264 if (!test_bit(R5_ReWrite
, &dev
->flags
)) {
5265 set_bit(R5_Wantwrite
, &dev
->flags
);
5266 set_bit(R5_ReWrite
, &dev
->flags
);
5268 /* let's read it back */
5269 set_bit(R5_Wantread
, &dev
->flags
);
5270 set_bit(R5_LOCKED
, &dev
->flags
);
5275 /* Finish reconstruct operations initiated by the expansion process */
5276 if (sh
->reconstruct_state
== reconstruct_state_result
) {
5277 struct stripe_head
*sh_src
5278 = raid5_get_active_stripe(conf
, NULL
, sh
->sector
,
5279 R5_GAS_PREVIOUS
| R5_GAS_NOBLOCK
|
5281 if (sh_src
&& test_bit(STRIPE_EXPAND_SOURCE
, &sh_src
->state
)) {
5282 /* sh cannot be written until sh_src has been read.
5283 * so arrange for sh to be delayed a little
5285 set_bit(STRIPE_DELAYED
, &sh
->state
);
5286 set_bit(STRIPE_HANDLE
, &sh
->state
);
5287 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE
,
5289 atomic_inc(&conf
->preread_active_stripes
);
5290 raid5_release_stripe(sh_src
);
5294 raid5_release_stripe(sh_src
);
5296 sh
->reconstruct_state
= reconstruct_state_idle
;
5297 clear_bit(STRIPE_EXPANDING
, &sh
->state
);
5298 for (i
= conf
->raid_disks
; i
--; ) {
5299 set_bit(R5_Wantwrite
, &sh
->dev
[i
].flags
);
5300 set_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
5305 if (s
.expanded
&& test_bit(STRIPE_EXPANDING
, &sh
->state
) &&
5306 !sh
->reconstruct_state
) {
5307 /* Need to write out all blocks after computing parity */
5308 sh
->disks
= conf
->raid_disks
;
5309 stripe_set_idx(sh
->sector
, conf
, 0, sh
);
5310 schedule_reconstruction(sh
, &s
, 1, 1);
5311 } else if (s
.expanded
&& !sh
->reconstruct_state
&& s
.locked
== 0) {
5312 clear_bit(STRIPE_EXPAND_READY
, &sh
->state
);
5313 atomic_dec(&conf
->reshape_stripes
);
5314 wake_up(&conf
->wait_for_overlap
);
5315 md_done_sync(conf
->mddev
, RAID5_STRIPE_SECTORS(conf
), 1);
5318 if (s
.expanding
&& s
.locked
== 0 &&
5319 !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
))
5320 handle_stripe_expansion(conf
, sh
);
5323 /* wait for this device to become unblocked */
5324 if (unlikely(s
.blocked_rdev
)) {
5325 if (conf
->mddev
->external
)
5326 md_wait_for_blocked_rdev(s
.blocked_rdev
,
5329 /* Internal metadata will immediately
5330 * be written by raid5d, so we don't
5331 * need to wait here.
5333 rdev_dec_pending(s
.blocked_rdev
,
5337 if (s
.handle_bad_blocks
)
5338 for (i
= disks
; i
--; ) {
5339 struct md_rdev
*rdev
;
5340 struct r5dev
*dev
= &sh
->dev
[i
];
5341 if (test_and_clear_bit(R5_WriteError
, &dev
->flags
)) {
5342 /* We own a safe reference to the rdev */
5343 rdev
= rdev_pend_deref(conf
->disks
[i
].rdev
);
5344 if (!rdev_set_badblocks(rdev
, sh
->sector
,
5345 RAID5_STRIPE_SECTORS(conf
), 0))
5346 md_error(conf
->mddev
, rdev
);
5347 rdev_dec_pending(rdev
, conf
->mddev
);
5349 if (test_and_clear_bit(R5_MadeGood
, &dev
->flags
)) {
5350 rdev
= rdev_pend_deref(conf
->disks
[i
].rdev
);
5351 rdev_clear_badblocks(rdev
, sh
->sector
,
5352 RAID5_STRIPE_SECTORS(conf
), 0);
5353 rdev_dec_pending(rdev
, conf
->mddev
);
5355 if (test_and_clear_bit(R5_MadeGoodRepl
, &dev
->flags
)) {
5356 rdev
= rdev_pend_deref(conf
->disks
[i
].replacement
);
5358 /* rdev have been moved down */
5359 rdev
= rdev_pend_deref(conf
->disks
[i
].rdev
);
5360 rdev_clear_badblocks(rdev
, sh
->sector
,
5361 RAID5_STRIPE_SECTORS(conf
), 0);
5362 rdev_dec_pending(rdev
, conf
->mddev
);
5367 raid_run_ops(sh
, s
.ops_request
);
5371 if (s
.dec_preread_active
) {
5372 /* We delay this until after ops_run_io so that if make_request
5373 * is waiting on a flush, it won't continue until the writes
5374 * have actually been submitted.
5376 atomic_dec(&conf
->preread_active_stripes
);
5377 if (atomic_read(&conf
->preread_active_stripes
) <
5379 md_wakeup_thread(conf
->mddev
->thread
);
5382 clear_bit_unlock(STRIPE_ACTIVE
, &sh
->state
);
5385 static void raid5_activate_delayed(struct r5conf
*conf
)
5386 __must_hold(&conf
->device_lock
)
5388 if (atomic_read(&conf
->preread_active_stripes
) < IO_THRESHOLD
) {
5389 while (!list_empty(&conf
->delayed_list
)) {
5390 struct list_head
*l
= conf
->delayed_list
.next
;
5391 struct stripe_head
*sh
;
5392 sh
= list_entry(l
, struct stripe_head
, lru
);
5394 clear_bit(STRIPE_DELAYED
, &sh
->state
);
5395 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
5396 atomic_inc(&conf
->preread_active_stripes
);
5397 list_add_tail(&sh
->lru
, &conf
->hold_list
);
5398 raid5_wakeup_stripe_thread(sh
);
5403 static void activate_bit_delay(struct r5conf
*conf
,
5404 struct list_head
*temp_inactive_list
)
5405 __must_hold(&conf
->device_lock
)
5407 struct list_head head
;
5408 list_add(&head
, &conf
->bitmap_list
);
5409 list_del_init(&conf
->bitmap_list
);
5410 while (!list_empty(&head
)) {
5411 struct stripe_head
*sh
= list_entry(head
.next
, struct stripe_head
, lru
);
5413 list_del_init(&sh
->lru
);
5414 atomic_inc(&sh
->count
);
5415 hash
= sh
->hash_lock_index
;
5416 __release_stripe(conf
, sh
, &temp_inactive_list
[hash
]);
5420 static int in_chunk_boundary(struct mddev
*mddev
, struct bio
*bio
)
5422 struct r5conf
*conf
= mddev
->private;
5423 sector_t sector
= bio
->bi_iter
.bi_sector
;
5424 unsigned int chunk_sectors
;
5425 unsigned int bio_sectors
= bio_sectors(bio
);
5427 chunk_sectors
= min(conf
->chunk_sectors
, conf
->prev_chunk_sectors
);
5428 return chunk_sectors
>=
5429 ((sector
& (chunk_sectors
- 1)) + bio_sectors
);
5433 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
5434 * later sampled by raid5d.
5436 static void add_bio_to_retry(struct bio
*bi
,struct r5conf
*conf
)
5438 unsigned long flags
;
5440 spin_lock_irqsave(&conf
->device_lock
, flags
);
5442 bi
->bi_next
= conf
->retry_read_aligned_list
;
5443 conf
->retry_read_aligned_list
= bi
;
5445 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
5446 md_wakeup_thread(conf
->mddev
->thread
);
5449 static struct bio
*remove_bio_from_retry(struct r5conf
*conf
,
5450 unsigned int *offset
)
5454 bi
= conf
->retry_read_aligned
;
5456 *offset
= conf
->retry_read_offset
;
5457 conf
->retry_read_aligned
= NULL
;
5460 bi
= conf
->retry_read_aligned_list
;
5462 conf
->retry_read_aligned_list
= bi
->bi_next
;
5471 * The "raid5_align_endio" should check if the read succeeded and if it
5472 * did, call bio_endio on the original bio (having bio_put the new bio
5474 * If the read failed..
5476 static void raid5_align_endio(struct bio
*bi
)
5478 struct bio
*raid_bi
= bi
->bi_private
;
5479 struct md_rdev
*rdev
= (void *)raid_bi
->bi_next
;
5480 struct mddev
*mddev
= rdev
->mddev
;
5481 struct r5conf
*conf
= mddev
->private;
5482 blk_status_t error
= bi
->bi_status
;
5485 raid_bi
->bi_next
= NULL
;
5486 rdev_dec_pending(rdev
, conf
->mddev
);
5490 if (atomic_dec_and_test(&conf
->active_aligned_reads
))
5491 wake_up(&conf
->wait_for_quiescent
);
5495 pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
5497 add_bio_to_retry(raid_bi
, conf
);
5500 static int raid5_read_one_chunk(struct mddev
*mddev
, struct bio
*raid_bio
)
5502 struct r5conf
*conf
= mddev
->private;
5503 struct bio
*align_bio
;
5504 struct md_rdev
*rdev
;
5505 sector_t sector
, end_sector
, first_bad
;
5506 int bad_sectors
, dd_idx
;
5509 if (!in_chunk_boundary(mddev
, raid_bio
)) {
5510 pr_debug("%s: non aligned\n", __func__
);
5514 sector
= raid5_compute_sector(conf
, raid_bio
->bi_iter
.bi_sector
, 0,
5516 end_sector
= sector
+ bio_sectors(raid_bio
);
5519 if (r5c_big_stripe_cached(conf
, sector
))
5520 goto out_rcu_unlock
;
5522 rdev
= rcu_dereference(conf
->disks
[dd_idx
].replacement
);
5523 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
) ||
5524 rdev
->recovery_offset
< end_sector
) {
5525 rdev
= rcu_dereference(conf
->disks
[dd_idx
].rdev
);
5527 goto out_rcu_unlock
;
5528 if (test_bit(Faulty
, &rdev
->flags
) ||
5529 !(test_bit(In_sync
, &rdev
->flags
) ||
5530 rdev
->recovery_offset
>= end_sector
))
5531 goto out_rcu_unlock
;
5534 atomic_inc(&rdev
->nr_pending
);
5537 if (is_badblock(rdev
, sector
, bio_sectors(raid_bio
), &first_bad
,
5539 rdev_dec_pending(rdev
, mddev
);
5543 md_account_bio(mddev
, &raid_bio
);
5544 raid_bio
->bi_next
= (void *)rdev
;
5546 align_bio
= bio_alloc_clone(rdev
->bdev
, raid_bio
, GFP_NOIO
,
5548 align_bio
->bi_end_io
= raid5_align_endio
;
5549 align_bio
->bi_private
= raid_bio
;
5550 align_bio
->bi_iter
.bi_sector
= sector
;
5552 /* No reshape active, so we can trust rdev->data_offset */
5553 align_bio
->bi_iter
.bi_sector
+= rdev
->data_offset
;
5556 if (conf
->quiesce
== 0) {
5557 atomic_inc(&conf
->active_aligned_reads
);
5560 /* need a memory barrier to detect the race with raid5_quiesce() */
5561 if (!did_inc
|| smp_load_acquire(&conf
->quiesce
) != 0) {
5562 /* quiesce is in progress, so we need to undo io activation and wait
5565 if (did_inc
&& atomic_dec_and_test(&conf
->active_aligned_reads
))
5566 wake_up(&conf
->wait_for_quiescent
);
5567 spin_lock_irq(&conf
->device_lock
);
5568 wait_event_lock_irq(conf
->wait_for_quiescent
, conf
->quiesce
== 0,
5570 atomic_inc(&conf
->active_aligned_reads
);
5571 spin_unlock_irq(&conf
->device_lock
);
5575 trace_block_bio_remap(align_bio
, disk_devt(mddev
->gendisk
),
5576 raid_bio
->bi_iter
.bi_sector
);
5577 submit_bio_noacct(align_bio
);
5585 static struct bio
*chunk_aligned_read(struct mddev
*mddev
, struct bio
*raid_bio
)
5588 sector_t sector
= raid_bio
->bi_iter
.bi_sector
;
5589 unsigned chunk_sects
= mddev
->chunk_sectors
;
5590 unsigned sectors
= chunk_sects
- (sector
& (chunk_sects
-1));
5592 if (sectors
< bio_sectors(raid_bio
)) {
5593 struct r5conf
*conf
= mddev
->private;
5594 split
= bio_split(raid_bio
, sectors
, GFP_NOIO
, &conf
->bio_split
);
5595 bio_chain(split
, raid_bio
);
5596 submit_bio_noacct(raid_bio
);
5600 if (!raid5_read_one_chunk(mddev
, raid_bio
))
5606 /* __get_priority_stripe - get the next stripe to process
5608 * Full stripe writes are allowed to pass preread active stripes up until
5609 * the bypass_threshold is exceeded. In general the bypass_count
5610 * increments when the handle_list is handled before the hold_list; however, it
5611 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
5612 * stripe with in flight i/o. The bypass_count will be reset when the
5613 * head of the hold_list has changed, i.e. the head was promoted to the
5616 static struct stripe_head
*__get_priority_stripe(struct r5conf
*conf
, int group
)
5617 __must_hold(&conf
->device_lock
)
5619 struct stripe_head
*sh
, *tmp
;
5620 struct list_head
*handle_list
= NULL
;
5621 struct r5worker_group
*wg
;
5622 bool second_try
= !r5c_is_writeback(conf
->log
) &&
5623 !r5l_log_disk_error(conf
);
5624 bool try_loprio
= test_bit(R5C_LOG_TIGHT
, &conf
->cache_state
) ||
5625 r5l_log_disk_error(conf
);
5630 if (conf
->worker_cnt_per_group
== 0) {
5631 handle_list
= try_loprio
? &conf
->loprio_list
:
5633 } else if (group
!= ANY_GROUP
) {
5634 handle_list
= try_loprio
? &conf
->worker_groups
[group
].loprio_list
:
5635 &conf
->worker_groups
[group
].handle_list
;
5636 wg
= &conf
->worker_groups
[group
];
5639 for (i
= 0; i
< conf
->group_cnt
; i
++) {
5640 handle_list
= try_loprio
? &conf
->worker_groups
[i
].loprio_list
:
5641 &conf
->worker_groups
[i
].handle_list
;
5642 wg
= &conf
->worker_groups
[i
];
5643 if (!list_empty(handle_list
))
5648 pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
5650 list_empty(handle_list
) ? "empty" : "busy",
5651 list_empty(&conf
->hold_list
) ? "empty" : "busy",
5652 atomic_read(&conf
->pending_full_writes
), conf
->bypass_count
);
5654 if (!list_empty(handle_list
)) {
5655 sh
= list_entry(handle_list
->next
, typeof(*sh
), lru
);
5657 if (list_empty(&conf
->hold_list
))
5658 conf
->bypass_count
= 0;
5659 else if (!test_bit(STRIPE_IO_STARTED
, &sh
->state
)) {
5660 if (conf
->hold_list
.next
== conf
->last_hold
)
5661 conf
->bypass_count
++;
5663 conf
->last_hold
= conf
->hold_list
.next
;
5664 conf
->bypass_count
-= conf
->bypass_threshold
;
5665 if (conf
->bypass_count
< 0)
5666 conf
->bypass_count
= 0;
5669 } else if (!list_empty(&conf
->hold_list
) &&
5670 ((conf
->bypass_threshold
&&
5671 conf
->bypass_count
> conf
->bypass_threshold
) ||
5672 atomic_read(&conf
->pending_full_writes
) == 0)) {
5674 list_for_each_entry(tmp
, &conf
->hold_list
, lru
) {
5675 if (conf
->worker_cnt_per_group
== 0 ||
5676 group
== ANY_GROUP
||
5677 !cpu_online(tmp
->cpu
) ||
5678 cpu_to_group(tmp
->cpu
) == group
) {
5685 conf
->bypass_count
-= conf
->bypass_threshold
;
5686 if (conf
->bypass_count
< 0)
5687 conf
->bypass_count
= 0;
5696 try_loprio
= !try_loprio
;
5704 list_del_init(&sh
->lru
);
5705 BUG_ON(atomic_inc_return(&sh
->count
) != 1);
5709 struct raid5_plug_cb
{
5710 struct blk_plug_cb cb
;
5711 struct list_head list
;
5712 struct list_head temp_inactive_list
[NR_STRIPE_HASH_LOCKS
];
5715 static void raid5_unplug(struct blk_plug_cb
*blk_cb
, bool from_schedule
)
5717 struct raid5_plug_cb
*cb
= container_of(
5718 blk_cb
, struct raid5_plug_cb
, cb
);
5719 struct stripe_head
*sh
;
5720 struct mddev
*mddev
= cb
->cb
.data
;
5721 struct r5conf
*conf
= mddev
->private;
5725 if (cb
->list
.next
&& !list_empty(&cb
->list
)) {
5726 spin_lock_irq(&conf
->device_lock
);
5727 while (!list_empty(&cb
->list
)) {
5728 sh
= list_first_entry(&cb
->list
, struct stripe_head
, lru
);
5729 list_del_init(&sh
->lru
);
5731 * avoid race release_stripe_plug() sees
5732 * STRIPE_ON_UNPLUG_LIST clear but the stripe
5733 * is still in our list
5735 smp_mb__before_atomic();
5736 clear_bit(STRIPE_ON_UNPLUG_LIST
, &sh
->state
);
5738 * STRIPE_ON_RELEASE_LIST could be set here. In that
5739 * case, the count is always > 1 here
5741 hash
= sh
->hash_lock_index
;
5742 __release_stripe(conf
, sh
, &cb
->temp_inactive_list
[hash
]);
5745 spin_unlock_irq(&conf
->device_lock
);
5747 release_inactive_stripe_list(conf
, cb
->temp_inactive_list
,
5748 NR_STRIPE_HASH_LOCKS
);
5750 trace_block_unplug(mddev
->queue
, cnt
, !from_schedule
);
5754 static void release_stripe_plug(struct mddev
*mddev
,
5755 struct stripe_head
*sh
)
5757 struct blk_plug_cb
*blk_cb
= blk_check_plugged(
5758 raid5_unplug
, mddev
,
5759 sizeof(struct raid5_plug_cb
));
5760 struct raid5_plug_cb
*cb
;
5763 raid5_release_stripe(sh
);
5767 cb
= container_of(blk_cb
, struct raid5_plug_cb
, cb
);
5769 if (cb
->list
.next
== NULL
) {
5771 INIT_LIST_HEAD(&cb
->list
);
5772 for (i
= 0; i
< NR_STRIPE_HASH_LOCKS
; i
++)
5773 INIT_LIST_HEAD(cb
->temp_inactive_list
+ i
);
5776 if (!test_and_set_bit(STRIPE_ON_UNPLUG_LIST
, &sh
->state
))
5777 list_add_tail(&sh
->lru
, &cb
->list
);
5779 raid5_release_stripe(sh
);
5782 static void make_discard_request(struct mddev
*mddev
, struct bio
*bi
)
5784 struct r5conf
*conf
= mddev
->private;
5785 sector_t logical_sector
, last_sector
;
5786 struct stripe_head
*sh
;
5789 /* We need to handle this when io_uring supports discard/trim */
5790 if (WARN_ON_ONCE(bi
->bi_opf
& REQ_NOWAIT
))
5793 if (mddev
->reshape_position
!= MaxSector
)
5794 /* Skip discard while reshape is happening */
5797 logical_sector
= bi
->bi_iter
.bi_sector
& ~((sector_t
)RAID5_STRIPE_SECTORS(conf
)-1);
5798 last_sector
= bio_end_sector(bi
);
5802 stripe_sectors
= conf
->chunk_sectors
*
5803 (conf
->raid_disks
- conf
->max_degraded
);
5804 logical_sector
= DIV_ROUND_UP_SECTOR_T(logical_sector
,
5806 sector_div(last_sector
, stripe_sectors
);
5808 logical_sector
*= conf
->chunk_sectors
;
5809 last_sector
*= conf
->chunk_sectors
;
5811 for (; logical_sector
< last_sector
;
5812 logical_sector
+= RAID5_STRIPE_SECTORS(conf
)) {
5816 sh
= raid5_get_active_stripe(conf
, NULL
, logical_sector
, 0);
5817 prepare_to_wait(&conf
->wait_for_overlap
, &w
,
5818 TASK_UNINTERRUPTIBLE
);
5819 set_bit(R5_Overlap
, &sh
->dev
[sh
->pd_idx
].flags
);
5820 if (test_bit(STRIPE_SYNCING
, &sh
->state
)) {
5821 raid5_release_stripe(sh
);
5825 clear_bit(R5_Overlap
, &sh
->dev
[sh
->pd_idx
].flags
);
5826 spin_lock_irq(&sh
->stripe_lock
);
5827 for (d
= 0; d
< conf
->raid_disks
; d
++) {
5828 if (d
== sh
->pd_idx
|| d
== sh
->qd_idx
)
5830 if (sh
->dev
[d
].towrite
|| sh
->dev
[d
].toread
) {
5831 set_bit(R5_Overlap
, &sh
->dev
[d
].flags
);
5832 spin_unlock_irq(&sh
->stripe_lock
);
5833 raid5_release_stripe(sh
);
5838 set_bit(STRIPE_DISCARD
, &sh
->state
);
5839 finish_wait(&conf
->wait_for_overlap
, &w
);
5840 sh
->overwrite_disks
= 0;
5841 for (d
= 0; d
< conf
->raid_disks
; d
++) {
5842 if (d
== sh
->pd_idx
|| d
== sh
->qd_idx
)
5844 sh
->dev
[d
].towrite
= bi
;
5845 set_bit(R5_OVERWRITE
, &sh
->dev
[d
].flags
);
5846 bio_inc_remaining(bi
);
5847 md_write_inc(mddev
, bi
);
5848 sh
->overwrite_disks
++;
5850 spin_unlock_irq(&sh
->stripe_lock
);
5851 if (conf
->mddev
->bitmap
) {
5853 d
< conf
->raid_disks
- conf
->max_degraded
;
5855 md_bitmap_startwrite(mddev
->bitmap
,
5857 RAID5_STRIPE_SECTORS(conf
),
5859 sh
->bm_seq
= conf
->seq_flush
+ 1;
5860 set_bit(STRIPE_BIT_DELAY
, &sh
->state
);
5863 set_bit(STRIPE_HANDLE
, &sh
->state
);
5864 clear_bit(STRIPE_DELAYED
, &sh
->state
);
5865 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
5866 atomic_inc(&conf
->preread_active_stripes
);
5867 release_stripe_plug(mddev
, sh
);
5873 static bool ahead_of_reshape(struct mddev
*mddev
, sector_t sector
,
5874 sector_t reshape_sector
)
5876 return mddev
->reshape_backwards
? sector
< reshape_sector
:
5877 sector
>= reshape_sector
;
5880 static bool range_ahead_of_reshape(struct mddev
*mddev
, sector_t min
,
5881 sector_t max
, sector_t reshape_sector
)
5883 return mddev
->reshape_backwards
? max
< reshape_sector
:
5884 min
>= reshape_sector
;
5887 static bool stripe_ahead_of_reshape(struct mddev
*mddev
, struct r5conf
*conf
,
5888 struct stripe_head
*sh
)
5890 sector_t max_sector
= 0, min_sector
= MaxSector
;
5894 for (dd_idx
= 0; dd_idx
< sh
->disks
; dd_idx
++) {
5895 if (dd_idx
== sh
->pd_idx
)
5898 min_sector
= min(min_sector
, sh
->dev
[dd_idx
].sector
);
5899 max_sector
= min(max_sector
, sh
->dev
[dd_idx
].sector
);
5902 spin_lock_irq(&conf
->device_lock
);
5904 if (!range_ahead_of_reshape(mddev
, min_sector
, max_sector
,
5905 conf
->reshape_progress
))
5906 /* mismatch, need to try again */
5909 spin_unlock_irq(&conf
->device_lock
);
5914 static int add_all_stripe_bios(struct r5conf
*conf
,
5915 struct stripe_request_ctx
*ctx
, struct stripe_head
*sh
,
5916 struct bio
*bi
, int forwrite
, int previous
)
5921 spin_lock_irq(&sh
->stripe_lock
);
5923 for (dd_idx
= 0; dd_idx
< sh
->disks
; dd_idx
++) {
5924 struct r5dev
*dev
= &sh
->dev
[dd_idx
];
5926 if (dd_idx
== sh
->pd_idx
|| dd_idx
== sh
->qd_idx
)
5929 if (dev
->sector
< ctx
->first_sector
||
5930 dev
->sector
>= ctx
->last_sector
)
5933 if (stripe_bio_overlaps(sh
, bi
, dd_idx
, forwrite
)) {
5934 set_bit(R5_Overlap
, &dev
->flags
);
5943 for (dd_idx
= 0; dd_idx
< sh
->disks
; dd_idx
++) {
5944 struct r5dev
*dev
= &sh
->dev
[dd_idx
];
5946 if (dd_idx
== sh
->pd_idx
|| dd_idx
== sh
->qd_idx
)
5949 if (dev
->sector
< ctx
->first_sector
||
5950 dev
->sector
>= ctx
->last_sector
)
5953 __add_stripe_bio(sh
, bi
, dd_idx
, forwrite
, previous
);
5954 clear_bit((dev
->sector
- ctx
->first_sector
) >>
5955 RAID5_STRIPE_SHIFT(conf
), ctx
->sectors_to_do
);
5959 spin_unlock_irq(&sh
->stripe_lock
);
5963 static bool reshape_inprogress(struct mddev
*mddev
)
5965 return test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
) &&
5966 test_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
) &&
5967 !test_bit(MD_RECOVERY_DONE
, &mddev
->recovery
) &&
5968 !test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
);
5971 static bool reshape_disabled(struct mddev
*mddev
)
5973 return is_md_suspended(mddev
) || !md_is_rdwr(mddev
);
5976 static enum stripe_result
make_stripe_request(struct mddev
*mddev
,
5977 struct r5conf
*conf
, struct stripe_request_ctx
*ctx
,
5978 sector_t logical_sector
, struct bio
*bi
)
5980 const int rw
= bio_data_dir(bi
);
5981 enum stripe_result ret
;
5982 struct stripe_head
*sh
;
5983 sector_t new_sector
;
5984 int previous
= 0, flags
= 0;
5987 seq
= read_seqcount_begin(&conf
->gen_lock
);
5989 if (unlikely(conf
->reshape_progress
!= MaxSector
)) {
5991 * Spinlock is needed as reshape_progress may be
5992 * 64bit on a 32bit platform, and so it might be
5993 * possible to see a half-updated value
5994 * Of course reshape_progress could change after
5995 * the lock is dropped, so once we get a reference
5996 * to the stripe that we think it is, we will have
5999 spin_lock_irq(&conf
->device_lock
);
6000 if (ahead_of_reshape(mddev
, logical_sector
,
6001 conf
->reshape_progress
)) {
6004 if (ahead_of_reshape(mddev
, logical_sector
,
6005 conf
->reshape_safe
)) {
6006 spin_unlock_irq(&conf
->device_lock
);
6007 ret
= STRIPE_SCHEDULE_AND_RETRY
;
6011 spin_unlock_irq(&conf
->device_lock
);
6014 new_sector
= raid5_compute_sector(conf
, logical_sector
, previous
,
6016 pr_debug("raid456: %s, sector %llu logical %llu\n", __func__
,
6017 new_sector
, logical_sector
);
6020 flags
|= R5_GAS_PREVIOUS
;
6021 if (bi
->bi_opf
& REQ_RAHEAD
)
6022 flags
|= R5_GAS_NOBLOCK
;
6023 sh
= raid5_get_active_stripe(conf
, ctx
, new_sector
, flags
);
6024 if (unlikely(!sh
)) {
6025 /* cannot get stripe, just give-up */
6026 bi
->bi_status
= BLK_STS_IOERR
;
6030 if (unlikely(previous
) &&
6031 stripe_ahead_of_reshape(mddev
, conf
, sh
)) {
6033 * Expansion moved on while waiting for a stripe.
6034 * Expansion could still move past after this
6035 * test, but as we are holding a reference to
6036 * 'sh', we know that if that happens,
6037 * STRIPE_EXPANDING will get set and the expansion
6038 * won't proceed until we finish with the stripe.
6040 ret
= STRIPE_SCHEDULE_AND_RETRY
;
6044 if (read_seqcount_retry(&conf
->gen_lock
, seq
)) {
6045 /* Might have got the wrong stripe_head by accident */
6050 if (test_bit(STRIPE_EXPANDING
, &sh
->state
) ||
6051 !add_all_stripe_bios(conf
, ctx
, sh
, bi
, rw
, previous
)) {
6053 * Stripe is busy expanding or add failed due to
6054 * overlap. Flush everything and wait a while.
6056 md_wakeup_thread(mddev
->thread
);
6057 ret
= STRIPE_SCHEDULE_AND_RETRY
;
6061 if (stripe_can_batch(sh
)) {
6062 stripe_add_to_batch_list(conf
, sh
, ctx
->batch_last
);
6063 if (ctx
->batch_last
)
6064 raid5_release_stripe(ctx
->batch_last
);
6065 atomic_inc(&sh
->count
);
6066 ctx
->batch_last
= sh
;
6069 if (ctx
->do_flush
) {
6070 set_bit(STRIPE_R5C_PREFLUSH
, &sh
->state
);
6071 /* we only need flush for one stripe */
6072 ctx
->do_flush
= false;
6075 set_bit(STRIPE_HANDLE
, &sh
->state
);
6076 clear_bit(STRIPE_DELAYED
, &sh
->state
);
6077 if ((!sh
->batch_head
|| sh
== sh
->batch_head
) &&
6078 (bi
->bi_opf
& REQ_SYNC
) &&
6079 !test_and_set_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
6080 atomic_inc(&conf
->preread_active_stripes
);
6082 release_stripe_plug(mddev
, sh
);
6083 return STRIPE_SUCCESS
;
6086 raid5_release_stripe(sh
);
6088 if (ret
== STRIPE_SCHEDULE_AND_RETRY
&& !reshape_inprogress(mddev
) &&
6089 reshape_disabled(mddev
)) {
6090 bi
->bi_status
= BLK_STS_IOERR
;
6092 pr_err("md/raid456:%s: io failed across reshape position while reshape can't make progress.\n",
6100 * If the bio covers multiple data disks, find sector within the bio that has
6101 * the lowest chunk offset in the first chunk.
6103 static sector_t
raid5_bio_lowest_chunk_sector(struct r5conf
*conf
,
6106 int sectors_per_chunk
= conf
->chunk_sectors
;
6107 int raid_disks
= conf
->raid_disks
;
6109 struct stripe_head sh
;
6110 unsigned int chunk_offset
;
6111 sector_t r_sector
= bi
->bi_iter
.bi_sector
& ~((sector_t
)RAID5_STRIPE_SECTORS(conf
)-1);
6114 /* We pass in fake stripe_head to get back parity disk numbers */
6115 sector
= raid5_compute_sector(conf
, r_sector
, 0, &dd_idx
, &sh
);
6116 chunk_offset
= sector_div(sector
, sectors_per_chunk
);
6117 if (sectors_per_chunk
- chunk_offset
>= bio_sectors(bi
))
6120 * Bio crosses to the next data disk. Check whether it's in the same
6124 while (dd_idx
== sh
.pd_idx
|| dd_idx
== sh
.qd_idx
)
6126 if (dd_idx
>= raid_disks
)
6128 return r_sector
+ sectors_per_chunk
- chunk_offset
;
6131 static bool raid5_make_request(struct mddev
*mddev
, struct bio
* bi
)
6133 DEFINE_WAIT_FUNC(wait
, woken_wake_function
);
6134 struct r5conf
*conf
= mddev
->private;
6135 sector_t logical_sector
;
6136 struct stripe_request_ctx ctx
= {};
6137 const int rw
= bio_data_dir(bi
);
6138 enum stripe_result res
;
6141 if (unlikely(bi
->bi_opf
& REQ_PREFLUSH
)) {
6142 int ret
= log_handle_flush_request(conf
, bi
);
6146 if (ret
== -ENODEV
) {
6147 if (md_flush_request(mddev
, bi
))
6150 /* ret == -EAGAIN, fallback */
6152 * if r5l_handle_flush_request() didn't clear REQ_PREFLUSH,
6153 * we need to flush journal device
6155 ctx
.do_flush
= bi
->bi_opf
& REQ_PREFLUSH
;
6158 if (!md_write_start(mddev
, bi
))
6161 * If array is degraded, better not do chunk aligned read because
6162 * later we might have to read it again in order to reconstruct
6163 * data on failed drives.
6165 if (rw
== READ
&& mddev
->degraded
== 0 &&
6166 mddev
->reshape_position
== MaxSector
) {
6167 bi
= chunk_aligned_read(mddev
, bi
);
6172 if (unlikely(bio_op(bi
) == REQ_OP_DISCARD
)) {
6173 make_discard_request(mddev
, bi
);
6174 md_write_end(mddev
);
6178 logical_sector
= bi
->bi_iter
.bi_sector
& ~((sector_t
)RAID5_STRIPE_SECTORS(conf
)-1);
6179 ctx
.first_sector
= logical_sector
;
6180 ctx
.last_sector
= bio_end_sector(bi
);
6183 stripe_cnt
= DIV_ROUND_UP_SECTOR_T(ctx
.last_sector
- logical_sector
,
6184 RAID5_STRIPE_SECTORS(conf
));
6185 bitmap_set(ctx
.sectors_to_do
, 0, stripe_cnt
);
6187 pr_debug("raid456: %s, logical %llu to %llu\n", __func__
,
6188 bi
->bi_iter
.bi_sector
, ctx
.last_sector
);
6190 /* Bail out if conflicts with reshape and REQ_NOWAIT is set */
6191 if ((bi
->bi_opf
& REQ_NOWAIT
) &&
6192 (conf
->reshape_progress
!= MaxSector
) &&
6193 !ahead_of_reshape(mddev
, logical_sector
, conf
->reshape_progress
) &&
6194 ahead_of_reshape(mddev
, logical_sector
, conf
->reshape_safe
)) {
6195 bio_wouldblock_error(bi
);
6197 md_write_end(mddev
);
6200 md_account_bio(mddev
, &bi
);
6203 * Lets start with the stripe with the lowest chunk offset in the first
6204 * chunk. That has the best chances of creating IOs adjacent to
6205 * previous IOs in case of sequential IO and thus creates the most
6206 * sequential IO pattern. We don't bother with the optimization when
6207 * reshaping as the performance benefit is not worth the complexity.
6209 if (likely(conf
->reshape_progress
== MaxSector
))
6210 logical_sector
= raid5_bio_lowest_chunk_sector(conf
, bi
);
6211 s
= (logical_sector
- ctx
.first_sector
) >> RAID5_STRIPE_SHIFT(conf
);
6213 add_wait_queue(&conf
->wait_for_overlap
, &wait
);
6215 res
= make_stripe_request(mddev
, conf
, &ctx
, logical_sector
,
6217 if (res
== STRIPE_FAIL
)
6220 if (res
== STRIPE_RETRY
)
6223 if (res
== STRIPE_SCHEDULE_AND_RETRY
) {
6225 * Must release the reference to batch_last before
6226 * scheduling and waiting for work to be done,
6227 * otherwise the batch_last stripe head could prevent
6228 * raid5_activate_delayed() from making progress
6229 * and thus deadlocking.
6231 if (ctx
.batch_last
) {
6232 raid5_release_stripe(ctx
.batch_last
);
6233 ctx
.batch_last
= NULL
;
6236 wait_woken(&wait
, TASK_UNINTERRUPTIBLE
,
6237 MAX_SCHEDULE_TIMEOUT
);
6241 s
= find_next_bit_wrap(ctx
.sectors_to_do
, stripe_cnt
, s
);
6242 if (s
== stripe_cnt
)
6245 logical_sector
= ctx
.first_sector
+
6246 (s
<< RAID5_STRIPE_SHIFT(conf
));
6248 remove_wait_queue(&conf
->wait_for_overlap
, &wait
);
6251 raid5_release_stripe(ctx
.batch_last
);
6254 md_write_end(mddev
);
6259 static sector_t
raid5_size(struct mddev
*mddev
, sector_t sectors
, int raid_disks
);
6261 static sector_t
reshape_request(struct mddev
*mddev
, sector_t sector_nr
, int *skipped
)
6263 /* reshaping is quite different to recovery/resync so it is
6264 * handled quite separately ... here.
6266 * On each call to sync_request, we gather one chunk worth of
6267 * destination stripes and flag them as expanding.
6268 * Then we find all the source stripes and request reads.
6269 * As the reads complete, handle_stripe will copy the data
6270 * into the destination stripe and release that stripe.
6272 struct r5conf
*conf
= mddev
->private;
6273 struct stripe_head
*sh
;
6274 struct md_rdev
*rdev
;
6275 sector_t first_sector
, last_sector
;
6276 int raid_disks
= conf
->previous_raid_disks
;
6277 int data_disks
= raid_disks
- conf
->max_degraded
;
6278 int new_data_disks
= conf
->raid_disks
- conf
->max_degraded
;
6281 sector_t writepos
, readpos
, safepos
;
6282 sector_t stripe_addr
;
6283 int reshape_sectors
;
6284 struct list_head stripes
;
6287 if (sector_nr
== 0) {
6288 /* If restarting in the middle, skip the initial sectors */
6289 if (mddev
->reshape_backwards
&&
6290 conf
->reshape_progress
< raid5_size(mddev
, 0, 0)) {
6291 sector_nr
= raid5_size(mddev
, 0, 0)
6292 - conf
->reshape_progress
;
6293 } else if (mddev
->reshape_backwards
&&
6294 conf
->reshape_progress
== MaxSector
) {
6295 /* shouldn't happen, but just in case, finish up.*/
6296 sector_nr
= MaxSector
;
6297 } else if (!mddev
->reshape_backwards
&&
6298 conf
->reshape_progress
> 0)
6299 sector_nr
= conf
->reshape_progress
;
6300 sector_div(sector_nr
, new_data_disks
);
6302 mddev
->curr_resync_completed
= sector_nr
;
6303 sysfs_notify_dirent_safe(mddev
->sysfs_completed
);
6310 /* We need to process a full chunk at a time.
6311 * If old and new chunk sizes differ, we need to process the
6315 reshape_sectors
= max(conf
->chunk_sectors
, conf
->prev_chunk_sectors
);
6317 /* We update the metadata at least every 10 seconds, or when
6318 * the data about to be copied would over-write the source of
6319 * the data at the front of the range. i.e. one new_stripe
6320 * along from reshape_progress new_maps to after where
6321 * reshape_safe old_maps to
6323 writepos
= conf
->reshape_progress
;
6324 sector_div(writepos
, new_data_disks
);
6325 readpos
= conf
->reshape_progress
;
6326 sector_div(readpos
, data_disks
);
6327 safepos
= conf
->reshape_safe
;
6328 sector_div(safepos
, data_disks
);
6329 if (mddev
->reshape_backwards
) {
6330 BUG_ON(writepos
< reshape_sectors
);
6331 writepos
-= reshape_sectors
;
6332 readpos
+= reshape_sectors
;
6333 safepos
+= reshape_sectors
;
6335 writepos
+= reshape_sectors
;
6336 /* readpos and safepos are worst-case calculations.
6337 * A negative number is overly pessimistic, and causes
6338 * obvious problems for unsigned storage. So clip to 0.
6340 readpos
-= min_t(sector_t
, reshape_sectors
, readpos
);
6341 safepos
-= min_t(sector_t
, reshape_sectors
, safepos
);
6344 /* Having calculated the 'writepos' possibly use it
6345 * to set 'stripe_addr' which is where we will write to.
6347 if (mddev
->reshape_backwards
) {
6348 BUG_ON(conf
->reshape_progress
== 0);
6349 stripe_addr
= writepos
;
6350 BUG_ON((mddev
->dev_sectors
&
6351 ~((sector_t
)reshape_sectors
- 1))
6352 - reshape_sectors
- stripe_addr
6355 BUG_ON(writepos
!= sector_nr
+ reshape_sectors
);
6356 stripe_addr
= sector_nr
;
6359 /* 'writepos' is the most advanced device address we might write.
6360 * 'readpos' is the least advanced device address we might read.
6361 * 'safepos' is the least address recorded in the metadata as having
6363 * If there is a min_offset_diff, these are adjusted either by
6364 * increasing the safepos/readpos if diff is negative, or
6365 * increasing writepos if diff is positive.
6366 * If 'readpos' is then behind 'writepos', there is no way that we can
6367 * ensure safety in the face of a crash - that must be done by userspace
6368 * making a backup of the data. So in that case there is no particular
6369 * rush to update metadata.
6370 * Otherwise if 'safepos' is behind 'writepos', then we really need to
6371 * update the metadata to advance 'safepos' to match 'readpos' so that
6372 * we can be safe in the event of a crash.
6373 * So we insist on updating metadata if safepos is behind writepos and
6374 * readpos is beyond writepos.
6375 * In any case, update the metadata every 10 seconds.
6376 * Maybe that number should be configurable, but I'm not sure it is
6377 * worth it.... maybe it could be a multiple of safemode_delay???
6379 if (conf
->min_offset_diff
< 0) {
6380 safepos
+= -conf
->min_offset_diff
;
6381 readpos
+= -conf
->min_offset_diff
;
6383 writepos
+= conf
->min_offset_diff
;
6385 if ((mddev
->reshape_backwards
6386 ? (safepos
> writepos
&& readpos
< writepos
)
6387 : (safepos
< writepos
&& readpos
> writepos
)) ||
6388 time_after(jiffies
, conf
->reshape_checkpoint
+ 10*HZ
)) {
6389 /* Cannot proceed until we've updated the superblock... */
6390 wait_event(conf
->wait_for_overlap
,
6391 atomic_read(&conf
->reshape_stripes
)==0
6392 || test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
));
6393 if (atomic_read(&conf
->reshape_stripes
) != 0)
6395 mddev
->reshape_position
= conf
->reshape_progress
;
6396 mddev
->curr_resync_completed
= sector_nr
;
6397 if (!mddev
->reshape_backwards
)
6398 /* Can update recovery_offset */
6399 rdev_for_each(rdev
, mddev
)
6400 if (rdev
->raid_disk
>= 0 &&
6401 !test_bit(Journal
, &rdev
->flags
) &&
6402 !test_bit(In_sync
, &rdev
->flags
) &&
6403 rdev
->recovery_offset
< sector_nr
)
6404 rdev
->recovery_offset
= sector_nr
;
6406 conf
->reshape_checkpoint
= jiffies
;
6407 set_bit(MD_SB_CHANGE_DEVS
, &mddev
->sb_flags
);
6408 md_wakeup_thread(mddev
->thread
);
6409 wait_event(mddev
->sb_wait
, mddev
->sb_flags
== 0 ||
6410 test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
));
6411 if (test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
))
6413 spin_lock_irq(&conf
->device_lock
);
6414 conf
->reshape_safe
= mddev
->reshape_position
;
6415 spin_unlock_irq(&conf
->device_lock
);
6416 wake_up(&conf
->wait_for_overlap
);
6417 sysfs_notify_dirent_safe(mddev
->sysfs_completed
);
6420 INIT_LIST_HEAD(&stripes
);
6421 for (i
= 0; i
< reshape_sectors
; i
+= RAID5_STRIPE_SECTORS(conf
)) {
6423 int skipped_disk
= 0;
6424 sh
= raid5_get_active_stripe(conf
, NULL
, stripe_addr
+i
,
6426 set_bit(STRIPE_EXPANDING
, &sh
->state
);
6427 atomic_inc(&conf
->reshape_stripes
);
6428 /* If any of this stripe is beyond the end of the old
6429 * array, then we need to zero those blocks
6431 for (j
=sh
->disks
; j
--;) {
6433 if (j
== sh
->pd_idx
)
6435 if (conf
->level
== 6 &&
6438 s
= raid5_compute_blocknr(sh
, j
, 0);
6439 if (s
< raid5_size(mddev
, 0, 0)) {
6443 memset(page_address(sh
->dev
[j
].page
), 0, RAID5_STRIPE_SIZE(conf
));
6444 set_bit(R5_Expanded
, &sh
->dev
[j
].flags
);
6445 set_bit(R5_UPTODATE
, &sh
->dev
[j
].flags
);
6447 if (!skipped_disk
) {
6448 set_bit(STRIPE_EXPAND_READY
, &sh
->state
);
6449 set_bit(STRIPE_HANDLE
, &sh
->state
);
6451 list_add(&sh
->lru
, &stripes
);
6453 spin_lock_irq(&conf
->device_lock
);
6454 if (mddev
->reshape_backwards
)
6455 conf
->reshape_progress
-= reshape_sectors
* new_data_disks
;
6457 conf
->reshape_progress
+= reshape_sectors
* new_data_disks
;
6458 spin_unlock_irq(&conf
->device_lock
);
6459 /* Ok, those stripe are ready. We can start scheduling
6460 * reads on the source stripes.
6461 * The source stripes are determined by mapping the first and last
6462 * block on the destination stripes.
6465 raid5_compute_sector(conf
, stripe_addr
*(new_data_disks
),
6468 raid5_compute_sector(conf
, ((stripe_addr
+reshape_sectors
)
6469 * new_data_disks
- 1),
6471 if (last_sector
>= mddev
->dev_sectors
)
6472 last_sector
= mddev
->dev_sectors
- 1;
6473 while (first_sector
<= last_sector
) {
6474 sh
= raid5_get_active_stripe(conf
, NULL
, first_sector
,
6475 R5_GAS_PREVIOUS
| R5_GAS_NOQUIESCE
);
6476 set_bit(STRIPE_EXPAND_SOURCE
, &sh
->state
);
6477 set_bit(STRIPE_HANDLE
, &sh
->state
);
6478 raid5_release_stripe(sh
);
6479 first_sector
+= RAID5_STRIPE_SECTORS(conf
);
6481 /* Now that the sources are clearly marked, we can release
6482 * the destination stripes
6484 while (!list_empty(&stripes
)) {
6485 sh
= list_entry(stripes
.next
, struct stripe_head
, lru
);
6486 list_del_init(&sh
->lru
);
6487 raid5_release_stripe(sh
);
6489 /* If this takes us to the resync_max point where we have to pause,
6490 * then we need to write out the superblock.
6492 sector_nr
+= reshape_sectors
;
6493 retn
= reshape_sectors
;
6495 if (mddev
->curr_resync_completed
> mddev
->resync_max
||
6496 (sector_nr
- mddev
->curr_resync_completed
) * 2
6497 >= mddev
->resync_max
- mddev
->curr_resync_completed
) {
6498 /* Cannot proceed until we've updated the superblock... */
6499 wait_event(conf
->wait_for_overlap
,
6500 atomic_read(&conf
->reshape_stripes
) == 0
6501 || test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
));
6502 if (atomic_read(&conf
->reshape_stripes
) != 0)
6504 mddev
->reshape_position
= conf
->reshape_progress
;
6505 mddev
->curr_resync_completed
= sector_nr
;
6506 if (!mddev
->reshape_backwards
)
6507 /* Can update recovery_offset */
6508 rdev_for_each(rdev
, mddev
)
6509 if (rdev
->raid_disk
>= 0 &&
6510 !test_bit(Journal
, &rdev
->flags
) &&
6511 !test_bit(In_sync
, &rdev
->flags
) &&
6512 rdev
->recovery_offset
< sector_nr
)
6513 rdev
->recovery_offset
= sector_nr
;
6514 conf
->reshape_checkpoint
= jiffies
;
6515 set_bit(MD_SB_CHANGE_DEVS
, &mddev
->sb_flags
);
6516 md_wakeup_thread(mddev
->thread
);
6517 wait_event(mddev
->sb_wait
,
6518 !test_bit(MD_SB_CHANGE_DEVS
, &mddev
->sb_flags
)
6519 || test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
));
6520 if (test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
))
6522 spin_lock_irq(&conf
->device_lock
);
6523 conf
->reshape_safe
= mddev
->reshape_position
;
6524 spin_unlock_irq(&conf
->device_lock
);
6525 wake_up(&conf
->wait_for_overlap
);
6526 sysfs_notify_dirent_safe(mddev
->sysfs_completed
);
6532 static inline sector_t
raid5_sync_request(struct mddev
*mddev
, sector_t sector_nr
,
6535 struct r5conf
*conf
= mddev
->private;
6536 struct stripe_head
*sh
;
6537 sector_t max_sector
= mddev
->dev_sectors
;
6538 sector_t sync_blocks
;
6539 int still_degraded
= 0;
6542 if (sector_nr
>= max_sector
) {
6543 /* just being told to finish up .. nothing much to do */
6545 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
)) {
6550 if (mddev
->curr_resync
< max_sector
) /* aborted */
6551 md_bitmap_end_sync(mddev
->bitmap
, mddev
->curr_resync
,
6553 else /* completed sync */
6555 md_bitmap_close_sync(mddev
->bitmap
);
6560 /* Allow raid5_quiesce to complete */
6561 wait_event(conf
->wait_for_overlap
, conf
->quiesce
!= 2);
6563 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
))
6564 return reshape_request(mddev
, sector_nr
, skipped
);
6566 /* No need to check resync_max as we never do more than one
6567 * stripe, and as resync_max will always be on a chunk boundary,
6568 * if the check in md_do_sync didn't fire, there is no chance
6569 * of overstepping resync_max here
6572 /* if there is too many failed drives and we are trying
6573 * to resync, then assert that we are finished, because there is
6574 * nothing we can do.
6576 if (mddev
->degraded
>= conf
->max_degraded
&&
6577 test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
)) {
6578 sector_t rv
= mddev
->dev_sectors
- sector_nr
;
6582 if (!test_bit(MD_RECOVERY_REQUESTED
, &mddev
->recovery
) &&
6584 !md_bitmap_start_sync(mddev
->bitmap
, sector_nr
, &sync_blocks
, 1) &&
6585 sync_blocks
>= RAID5_STRIPE_SECTORS(conf
)) {
6586 /* we can skip this block, and probably more */
6587 do_div(sync_blocks
, RAID5_STRIPE_SECTORS(conf
));
6589 /* keep things rounded to whole stripes */
6590 return sync_blocks
* RAID5_STRIPE_SECTORS(conf
);
6593 md_bitmap_cond_end_sync(mddev
->bitmap
, sector_nr
, false);
6595 sh
= raid5_get_active_stripe(conf
, NULL
, sector_nr
,
6598 sh
= raid5_get_active_stripe(conf
, NULL
, sector_nr
, 0);
6599 /* make sure we don't swamp the stripe cache if someone else
6600 * is trying to get access
6602 schedule_timeout_uninterruptible(1);
6604 /* Need to check if array will still be degraded after recovery/resync
6605 * Note in case of > 1 drive failures it's possible we're rebuilding
6606 * one drive while leaving another faulty drive in array.
6609 for (i
= 0; i
< conf
->raid_disks
; i
++) {
6610 struct md_rdev
*rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
6612 if (rdev
== NULL
|| test_bit(Faulty
, &rdev
->flags
))
6617 md_bitmap_start_sync(mddev
->bitmap
, sector_nr
, &sync_blocks
, still_degraded
);
6619 set_bit(STRIPE_SYNC_REQUESTED
, &sh
->state
);
6620 set_bit(STRIPE_HANDLE
, &sh
->state
);
6622 raid5_release_stripe(sh
);
6624 return RAID5_STRIPE_SECTORS(conf
);
6627 static int retry_aligned_read(struct r5conf
*conf
, struct bio
*raid_bio
,
6628 unsigned int offset
)
6630 /* We may not be able to submit a whole bio at once as there
6631 * may not be enough stripe_heads available.
6632 * We cannot pre-allocate enough stripe_heads as we may need
6633 * more than exist in the cache (if we allow ever large chunks).
6634 * So we do one stripe head at a time and record in
6635 * ->bi_hw_segments how many have been done.
6637 * We *know* that this entire raid_bio is in one chunk, so
6638 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
6640 struct stripe_head
*sh
;
6642 sector_t sector
, logical_sector
, last_sector
;
6646 logical_sector
= raid_bio
->bi_iter
.bi_sector
&
6647 ~((sector_t
)RAID5_STRIPE_SECTORS(conf
)-1);
6648 sector
= raid5_compute_sector(conf
, logical_sector
,
6650 last_sector
= bio_end_sector(raid_bio
);
6652 for (; logical_sector
< last_sector
;
6653 logical_sector
+= RAID5_STRIPE_SECTORS(conf
),
6654 sector
+= RAID5_STRIPE_SECTORS(conf
),
6658 /* already done this stripe */
6661 sh
= raid5_get_active_stripe(conf
, NULL
, sector
,
6662 R5_GAS_NOBLOCK
| R5_GAS_NOQUIESCE
);
6664 /* failed to get a stripe - must wait */
6665 conf
->retry_read_aligned
= raid_bio
;
6666 conf
->retry_read_offset
= scnt
;
6670 if (!add_stripe_bio(sh
, raid_bio
, dd_idx
, 0, 0)) {
6671 raid5_release_stripe(sh
);
6672 conf
->retry_read_aligned
= raid_bio
;
6673 conf
->retry_read_offset
= scnt
;
6677 set_bit(R5_ReadNoMerge
, &sh
->dev
[dd_idx
].flags
);
6679 raid5_release_stripe(sh
);
6683 bio_endio(raid_bio
);
6685 if (atomic_dec_and_test(&conf
->active_aligned_reads
))
6686 wake_up(&conf
->wait_for_quiescent
);
6690 static int handle_active_stripes(struct r5conf
*conf
, int group
,
6691 struct r5worker
*worker
,
6692 struct list_head
*temp_inactive_list
)
6693 __must_hold(&conf
->device_lock
)
6695 struct stripe_head
*batch
[MAX_STRIPE_BATCH
], *sh
;
6696 int i
, batch_size
= 0, hash
;
6697 bool release_inactive
= false;
6699 while (batch_size
< MAX_STRIPE_BATCH
&&
6700 (sh
= __get_priority_stripe(conf
, group
)) != NULL
)
6701 batch
[batch_size
++] = sh
;
6703 if (batch_size
== 0) {
6704 for (i
= 0; i
< NR_STRIPE_HASH_LOCKS
; i
++)
6705 if (!list_empty(temp_inactive_list
+ i
))
6707 if (i
== NR_STRIPE_HASH_LOCKS
) {
6708 spin_unlock_irq(&conf
->device_lock
);
6709 log_flush_stripe_to_raid(conf
);
6710 spin_lock_irq(&conf
->device_lock
);
6713 release_inactive
= true;
6715 spin_unlock_irq(&conf
->device_lock
);
6717 release_inactive_stripe_list(conf
, temp_inactive_list
,
6718 NR_STRIPE_HASH_LOCKS
);
6720 r5l_flush_stripe_to_raid(conf
->log
);
6721 if (release_inactive
) {
6722 spin_lock_irq(&conf
->device_lock
);
6726 for (i
= 0; i
< batch_size
; i
++)
6727 handle_stripe(batch
[i
]);
6728 log_write_stripe_run(conf
);
6732 spin_lock_irq(&conf
->device_lock
);
6733 for (i
= 0; i
< batch_size
; i
++) {
6734 hash
= batch
[i
]->hash_lock_index
;
6735 __release_stripe(conf
, batch
[i
], &temp_inactive_list
[hash
]);
6740 static void raid5_do_work(struct work_struct
*work
)
6742 struct r5worker
*worker
= container_of(work
, struct r5worker
, work
);
6743 struct r5worker_group
*group
= worker
->group
;
6744 struct r5conf
*conf
= group
->conf
;
6745 struct mddev
*mddev
= conf
->mddev
;
6746 int group_id
= group
- conf
->worker_groups
;
6748 struct blk_plug plug
;
6750 pr_debug("+++ raid5worker active\n");
6752 blk_start_plug(&plug
);
6754 spin_lock_irq(&conf
->device_lock
);
6756 int batch_size
, released
;
6758 released
= release_stripe_list(conf
, worker
->temp_inactive_list
);
6760 batch_size
= handle_active_stripes(conf
, group_id
, worker
,
6761 worker
->temp_inactive_list
);
6762 worker
->working
= false;
6763 if (!batch_size
&& !released
)
6765 handled
+= batch_size
;
6766 wait_event_lock_irq(mddev
->sb_wait
,
6767 !test_bit(MD_SB_CHANGE_PENDING
, &mddev
->sb_flags
),
6770 pr_debug("%d stripes handled\n", handled
);
6772 spin_unlock_irq(&conf
->device_lock
);
6774 flush_deferred_bios(conf
);
6776 r5l_flush_stripe_to_raid(conf
->log
);
6778 async_tx_issue_pending_all();
6779 blk_finish_plug(&plug
);
6781 pr_debug("--- raid5worker inactive\n");
6785 * This is our raid5 kernel thread.
6787 * We scan the hash table for stripes which can be handled now.
6788 * During the scan, completed stripes are saved for us by the interrupt
6789 * handler, so that they will not have to wait for our next wakeup.
6791 static void raid5d(struct md_thread
*thread
)
6793 struct mddev
*mddev
= thread
->mddev
;
6794 struct r5conf
*conf
= mddev
->private;
6796 struct blk_plug plug
;
6798 pr_debug("+++ raid5d active\n");
6800 md_check_recovery(mddev
);
6802 blk_start_plug(&plug
);
6804 spin_lock_irq(&conf
->device_lock
);
6807 int batch_size
, released
;
6808 unsigned int offset
;
6810 released
= release_stripe_list(conf
, conf
->temp_inactive_list
);
6812 clear_bit(R5_DID_ALLOC
, &conf
->cache_state
);
6815 !list_empty(&conf
->bitmap_list
)) {
6816 /* Now is a good time to flush some bitmap updates */
6818 spin_unlock_irq(&conf
->device_lock
);
6819 md_bitmap_unplug(mddev
->bitmap
);
6820 spin_lock_irq(&conf
->device_lock
);
6821 conf
->seq_write
= conf
->seq_flush
;
6822 activate_bit_delay(conf
, conf
->temp_inactive_list
);
6824 raid5_activate_delayed(conf
);
6826 while ((bio
= remove_bio_from_retry(conf
, &offset
))) {
6828 spin_unlock_irq(&conf
->device_lock
);
6829 ok
= retry_aligned_read(conf
, bio
, offset
);
6830 spin_lock_irq(&conf
->device_lock
);
6836 batch_size
= handle_active_stripes(conf
, ANY_GROUP
, NULL
,
6837 conf
->temp_inactive_list
);
6838 if (!batch_size
&& !released
)
6840 handled
+= batch_size
;
6842 if (mddev
->sb_flags
& ~(1 << MD_SB_CHANGE_PENDING
)) {
6843 spin_unlock_irq(&conf
->device_lock
);
6844 md_check_recovery(mddev
);
6845 spin_lock_irq(&conf
->device_lock
);
6848 * Waiting on MD_SB_CHANGE_PENDING below may deadlock
6849 * seeing md_check_recovery() is needed to clear
6850 * the flag when using mdmon.
6855 wait_event_lock_irq(mddev
->sb_wait
,
6856 !test_bit(MD_SB_CHANGE_PENDING
, &mddev
->sb_flags
),
6859 pr_debug("%d stripes handled\n", handled
);
6861 spin_unlock_irq(&conf
->device_lock
);
6862 if (test_and_clear_bit(R5_ALLOC_MORE
, &conf
->cache_state
) &&
6863 mutex_trylock(&conf
->cache_size_mutex
)) {
6864 grow_one_stripe(conf
, __GFP_NOWARN
);
6865 /* Set flag even if allocation failed. This helps
6866 * slow down allocation requests when mem is short
6868 set_bit(R5_DID_ALLOC
, &conf
->cache_state
);
6869 mutex_unlock(&conf
->cache_size_mutex
);
6872 flush_deferred_bios(conf
);
6874 r5l_flush_stripe_to_raid(conf
->log
);
6876 async_tx_issue_pending_all();
6877 blk_finish_plug(&plug
);
6879 pr_debug("--- raid5d inactive\n");
6883 raid5_show_stripe_cache_size(struct mddev
*mddev
, char *page
)
6885 struct r5conf
*conf
;
6887 spin_lock(&mddev
->lock
);
6888 conf
= mddev
->private;
6890 ret
= sprintf(page
, "%d\n", conf
->min_nr_stripes
);
6891 spin_unlock(&mddev
->lock
);
6896 raid5_set_cache_size(struct mddev
*mddev
, int size
)
6899 struct r5conf
*conf
= mddev
->private;
6901 if (size
<= 16 || size
> 32768)
6904 conf
->min_nr_stripes
= size
;
6905 mutex_lock(&conf
->cache_size_mutex
);
6906 while (size
< conf
->max_nr_stripes
&&
6907 drop_one_stripe(conf
))
6909 mutex_unlock(&conf
->cache_size_mutex
);
6911 md_allow_write(mddev
);
6913 mutex_lock(&conf
->cache_size_mutex
);
6914 while (size
> conf
->max_nr_stripes
)
6915 if (!grow_one_stripe(conf
, GFP_KERNEL
)) {
6916 conf
->min_nr_stripes
= conf
->max_nr_stripes
;
6920 mutex_unlock(&conf
->cache_size_mutex
);
6924 EXPORT_SYMBOL(raid5_set_cache_size
);
6927 raid5_store_stripe_cache_size(struct mddev
*mddev
, const char *page
, size_t len
)
6929 struct r5conf
*conf
;
6933 if (len
>= PAGE_SIZE
)
6935 if (kstrtoul(page
, 10, &new))
6937 err
= mddev_lock(mddev
);
6940 conf
= mddev
->private;
6944 err
= raid5_set_cache_size(mddev
, new);
6945 mddev_unlock(mddev
);
6950 static struct md_sysfs_entry
6951 raid5_stripecache_size
= __ATTR(stripe_cache_size
, S_IRUGO
| S_IWUSR
,
6952 raid5_show_stripe_cache_size
,
6953 raid5_store_stripe_cache_size
);
6956 raid5_show_rmw_level(struct mddev
*mddev
, char *page
)
6958 struct r5conf
*conf
= mddev
->private;
6960 return sprintf(page
, "%d\n", conf
->rmw_level
);
6966 raid5_store_rmw_level(struct mddev
*mddev
, const char *page
, size_t len
)
6968 struct r5conf
*conf
= mddev
->private;
6974 if (len
>= PAGE_SIZE
)
6977 if (kstrtoul(page
, 10, &new))
6980 if (new != PARITY_DISABLE_RMW
&& !raid6_call
.xor_syndrome
)
6983 if (new != PARITY_DISABLE_RMW
&&
6984 new != PARITY_ENABLE_RMW
&&
6985 new != PARITY_PREFER_RMW
)
6988 conf
->rmw_level
= new;
6992 static struct md_sysfs_entry
6993 raid5_rmw_level
= __ATTR(rmw_level
, S_IRUGO
| S_IWUSR
,
6994 raid5_show_rmw_level
,
6995 raid5_store_rmw_level
);
6998 raid5_show_stripe_size(struct mddev
*mddev
, char *page
)
7000 struct r5conf
*conf
;
7003 spin_lock(&mddev
->lock
);
7004 conf
= mddev
->private;
7006 ret
= sprintf(page
, "%lu\n", RAID5_STRIPE_SIZE(conf
));
7007 spin_unlock(&mddev
->lock
);
7011 #if PAGE_SIZE != DEFAULT_STRIPE_SIZE
7013 raid5_store_stripe_size(struct mddev
*mddev
, const char *page
, size_t len
)
7015 struct r5conf
*conf
;
7020 if (len
>= PAGE_SIZE
)
7022 if (kstrtoul(page
, 10, &new))
7026 * The value should not be bigger than PAGE_SIZE. It requires to
7027 * be multiple of DEFAULT_STRIPE_SIZE and the value should be power
7030 if (new % DEFAULT_STRIPE_SIZE
!= 0 ||
7031 new > PAGE_SIZE
|| new == 0 ||
7032 new != roundup_pow_of_two(new))
7035 err
= mddev_lock(mddev
);
7039 conf
= mddev
->private;
7045 if (new == conf
->stripe_size
)
7048 pr_debug("md/raid: change stripe_size from %lu to %lu\n",
7049 conf
->stripe_size
, new);
7051 if (mddev
->sync_thread
||
7052 test_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
) ||
7053 mddev
->reshape_position
!= MaxSector
||
7054 mddev
->sysfs_active
) {
7059 mddev_suspend(mddev
);
7060 mutex_lock(&conf
->cache_size_mutex
);
7061 size
= conf
->max_nr_stripes
;
7063 shrink_stripes(conf
);
7065 conf
->stripe_size
= new;
7066 conf
->stripe_shift
= ilog2(new) - 9;
7067 conf
->stripe_sectors
= new >> 9;
7068 if (grow_stripes(conf
, size
)) {
7069 pr_warn("md/raid:%s: couldn't allocate buffers\n",
7073 mutex_unlock(&conf
->cache_size_mutex
);
7074 mddev_resume(mddev
);
7077 mddev_unlock(mddev
);
7081 static struct md_sysfs_entry
7082 raid5_stripe_size
= __ATTR(stripe_size
, 0644,
7083 raid5_show_stripe_size
,
7084 raid5_store_stripe_size
);
7086 static struct md_sysfs_entry
7087 raid5_stripe_size
= __ATTR(stripe_size
, 0444,
7088 raid5_show_stripe_size
,
7093 raid5_show_preread_threshold(struct mddev
*mddev
, char *page
)
7095 struct r5conf
*conf
;
7097 spin_lock(&mddev
->lock
);
7098 conf
= mddev
->private;
7100 ret
= sprintf(page
, "%d\n", conf
->bypass_threshold
);
7101 spin_unlock(&mddev
->lock
);
7106 raid5_store_preread_threshold(struct mddev
*mddev
, const char *page
, size_t len
)
7108 struct r5conf
*conf
;
7112 if (len
>= PAGE_SIZE
)
7114 if (kstrtoul(page
, 10, &new))
7117 err
= mddev_lock(mddev
);
7120 conf
= mddev
->private;
7123 else if (new > conf
->min_nr_stripes
)
7126 conf
->bypass_threshold
= new;
7127 mddev_unlock(mddev
);
7131 static struct md_sysfs_entry
7132 raid5_preread_bypass_threshold
= __ATTR(preread_bypass_threshold
,
7134 raid5_show_preread_threshold
,
7135 raid5_store_preread_threshold
);
7138 raid5_show_skip_copy(struct mddev
*mddev
, char *page
)
7140 struct r5conf
*conf
;
7142 spin_lock(&mddev
->lock
);
7143 conf
= mddev
->private;
7145 ret
= sprintf(page
, "%d\n", conf
->skip_copy
);
7146 spin_unlock(&mddev
->lock
);
7151 raid5_store_skip_copy(struct mddev
*mddev
, const char *page
, size_t len
)
7153 struct r5conf
*conf
;
7157 if (len
>= PAGE_SIZE
)
7159 if (kstrtoul(page
, 10, &new))
7163 err
= mddev_lock(mddev
);
7166 conf
= mddev
->private;
7169 else if (new != conf
->skip_copy
) {
7170 struct request_queue
*q
= mddev
->queue
;
7172 mddev_suspend(mddev
);
7173 conf
->skip_copy
= new;
7175 blk_queue_flag_set(QUEUE_FLAG_STABLE_WRITES
, q
);
7177 blk_queue_flag_clear(QUEUE_FLAG_STABLE_WRITES
, q
);
7178 mddev_resume(mddev
);
7180 mddev_unlock(mddev
);
7184 static struct md_sysfs_entry
7185 raid5_skip_copy
= __ATTR(skip_copy
, S_IRUGO
| S_IWUSR
,
7186 raid5_show_skip_copy
,
7187 raid5_store_skip_copy
);
7190 stripe_cache_active_show(struct mddev
*mddev
, char *page
)
7192 struct r5conf
*conf
= mddev
->private;
7194 return sprintf(page
, "%d\n", atomic_read(&conf
->active_stripes
));
7199 static struct md_sysfs_entry
7200 raid5_stripecache_active
= __ATTR_RO(stripe_cache_active
);
7203 raid5_show_group_thread_cnt(struct mddev
*mddev
, char *page
)
7205 struct r5conf
*conf
;
7207 spin_lock(&mddev
->lock
);
7208 conf
= mddev
->private;
7210 ret
= sprintf(page
, "%d\n", conf
->worker_cnt_per_group
);
7211 spin_unlock(&mddev
->lock
);
7215 static int alloc_thread_groups(struct r5conf
*conf
, int cnt
,
7217 struct r5worker_group
**worker_groups
);
7219 raid5_store_group_thread_cnt(struct mddev
*mddev
, const char *page
, size_t len
)
7221 struct r5conf
*conf
;
7224 struct r5worker_group
*new_groups
, *old_groups
;
7227 if (len
>= PAGE_SIZE
)
7229 if (kstrtouint(page
, 10, &new))
7231 /* 8192 should be big enough */
7235 err
= mddev_lock(mddev
);
7238 conf
= mddev
->private;
7241 else if (new != conf
->worker_cnt_per_group
) {
7242 mddev_suspend(mddev
);
7244 old_groups
= conf
->worker_groups
;
7246 flush_workqueue(raid5_wq
);
7248 err
= alloc_thread_groups(conf
, new, &group_cnt
, &new_groups
);
7250 spin_lock_irq(&conf
->device_lock
);
7251 conf
->group_cnt
= group_cnt
;
7252 conf
->worker_cnt_per_group
= new;
7253 conf
->worker_groups
= new_groups
;
7254 spin_unlock_irq(&conf
->device_lock
);
7257 kfree(old_groups
[0].workers
);
7260 mddev_resume(mddev
);
7262 mddev_unlock(mddev
);
7267 static struct md_sysfs_entry
7268 raid5_group_thread_cnt
= __ATTR(group_thread_cnt
, S_IRUGO
| S_IWUSR
,
7269 raid5_show_group_thread_cnt
,
7270 raid5_store_group_thread_cnt
);
7272 static struct attribute
*raid5_attrs
[] = {
7273 &raid5_stripecache_size
.attr
,
7274 &raid5_stripecache_active
.attr
,
7275 &raid5_preread_bypass_threshold
.attr
,
7276 &raid5_group_thread_cnt
.attr
,
7277 &raid5_skip_copy
.attr
,
7278 &raid5_rmw_level
.attr
,
7279 &raid5_stripe_size
.attr
,
7280 &r5c_journal_mode
.attr
,
7281 &ppl_write_hint
.attr
,
7284 static const struct attribute_group raid5_attrs_group
= {
7286 .attrs
= raid5_attrs
,
7289 static int alloc_thread_groups(struct r5conf
*conf
, int cnt
, int *group_cnt
,
7290 struct r5worker_group
**worker_groups
)
7294 struct r5worker
*workers
;
7298 *worker_groups
= NULL
;
7301 *group_cnt
= num_possible_nodes();
7302 size
= sizeof(struct r5worker
) * cnt
;
7303 workers
= kcalloc(size
, *group_cnt
, GFP_NOIO
);
7304 *worker_groups
= kcalloc(*group_cnt
, sizeof(struct r5worker_group
),
7306 if (!*worker_groups
|| !workers
) {
7308 kfree(*worker_groups
);
7312 for (i
= 0; i
< *group_cnt
; i
++) {
7313 struct r5worker_group
*group
;
7315 group
= &(*worker_groups
)[i
];
7316 INIT_LIST_HEAD(&group
->handle_list
);
7317 INIT_LIST_HEAD(&group
->loprio_list
);
7319 group
->workers
= workers
+ i
* cnt
;
7321 for (j
= 0; j
< cnt
; j
++) {
7322 struct r5worker
*worker
= group
->workers
+ j
;
7323 worker
->group
= group
;
7324 INIT_WORK(&worker
->work
, raid5_do_work
);
7326 for (k
= 0; k
< NR_STRIPE_HASH_LOCKS
; k
++)
7327 INIT_LIST_HEAD(worker
->temp_inactive_list
+ k
);
7334 static void free_thread_groups(struct r5conf
*conf
)
7336 if (conf
->worker_groups
)
7337 kfree(conf
->worker_groups
[0].workers
);
7338 kfree(conf
->worker_groups
);
7339 conf
->worker_groups
= NULL
;
7343 raid5_size(struct mddev
*mddev
, sector_t sectors
, int raid_disks
)
7345 struct r5conf
*conf
= mddev
->private;
7348 sectors
= mddev
->dev_sectors
;
7350 /* size is defined by the smallest of previous and new size */
7351 raid_disks
= min(conf
->raid_disks
, conf
->previous_raid_disks
);
7353 sectors
&= ~((sector_t
)conf
->chunk_sectors
- 1);
7354 sectors
&= ~((sector_t
)conf
->prev_chunk_sectors
- 1);
7355 return sectors
* (raid_disks
- conf
->max_degraded
);
7358 static void free_scratch_buffer(struct r5conf
*conf
, struct raid5_percpu
*percpu
)
7360 safe_put_page(percpu
->spare_page
);
7361 percpu
->spare_page
= NULL
;
7362 kvfree(percpu
->scribble
);
7363 percpu
->scribble
= NULL
;
7366 static int alloc_scratch_buffer(struct r5conf
*conf
, struct raid5_percpu
*percpu
)
7368 if (conf
->level
== 6 && !percpu
->spare_page
) {
7369 percpu
->spare_page
= alloc_page(GFP_KERNEL
);
7370 if (!percpu
->spare_page
)
7374 if (scribble_alloc(percpu
,
7375 max(conf
->raid_disks
,
7376 conf
->previous_raid_disks
),
7377 max(conf
->chunk_sectors
,
7378 conf
->prev_chunk_sectors
)
7379 / RAID5_STRIPE_SECTORS(conf
))) {
7380 free_scratch_buffer(conf
, percpu
);
7384 local_lock_init(&percpu
->lock
);
7388 static int raid456_cpu_dead(unsigned int cpu
, struct hlist_node
*node
)
7390 struct r5conf
*conf
= hlist_entry_safe(node
, struct r5conf
, node
);
7392 free_scratch_buffer(conf
, per_cpu_ptr(conf
->percpu
, cpu
));
7396 static void raid5_free_percpu(struct r5conf
*conf
)
7401 cpuhp_state_remove_instance(CPUHP_MD_RAID5_PREPARE
, &conf
->node
);
7402 free_percpu(conf
->percpu
);
7405 static void free_conf(struct r5conf
*conf
)
7411 unregister_shrinker(&conf
->shrinker
);
7412 free_thread_groups(conf
);
7413 shrink_stripes(conf
);
7414 raid5_free_percpu(conf
);
7415 for (i
= 0; i
< conf
->pool_size
; i
++)
7416 if (conf
->disks
[i
].extra_page
)
7417 put_page(conf
->disks
[i
].extra_page
);
7419 bioset_exit(&conf
->bio_split
);
7420 kfree(conf
->stripe_hashtbl
);
7421 kfree(conf
->pending_data
);
7425 static int raid456_cpu_up_prepare(unsigned int cpu
, struct hlist_node
*node
)
7427 struct r5conf
*conf
= hlist_entry_safe(node
, struct r5conf
, node
);
7428 struct raid5_percpu
*percpu
= per_cpu_ptr(conf
->percpu
, cpu
);
7430 if (alloc_scratch_buffer(conf
, percpu
)) {
7431 pr_warn("%s: failed memory allocation for cpu%u\n",
7438 static int raid5_alloc_percpu(struct r5conf
*conf
)
7442 conf
->percpu
= alloc_percpu(struct raid5_percpu
);
7446 err
= cpuhp_state_add_instance(CPUHP_MD_RAID5_PREPARE
, &conf
->node
);
7448 conf
->scribble_disks
= max(conf
->raid_disks
,
7449 conf
->previous_raid_disks
);
7450 conf
->scribble_sectors
= max(conf
->chunk_sectors
,
7451 conf
->prev_chunk_sectors
);
7456 static unsigned long raid5_cache_scan(struct shrinker
*shrink
,
7457 struct shrink_control
*sc
)
7459 struct r5conf
*conf
= container_of(shrink
, struct r5conf
, shrinker
);
7460 unsigned long ret
= SHRINK_STOP
;
7462 if (mutex_trylock(&conf
->cache_size_mutex
)) {
7464 while (ret
< sc
->nr_to_scan
&&
7465 conf
->max_nr_stripes
> conf
->min_nr_stripes
) {
7466 if (drop_one_stripe(conf
) == 0) {
7472 mutex_unlock(&conf
->cache_size_mutex
);
7477 static unsigned long raid5_cache_count(struct shrinker
*shrink
,
7478 struct shrink_control
*sc
)
7480 struct r5conf
*conf
= container_of(shrink
, struct r5conf
, shrinker
);
7482 if (conf
->max_nr_stripes
< conf
->min_nr_stripes
)
7483 /* unlikely, but not impossible */
7485 return conf
->max_nr_stripes
- conf
->min_nr_stripes
;
7488 static struct r5conf
*setup_conf(struct mddev
*mddev
)
7490 struct r5conf
*conf
;
7491 int raid_disk
, memory
, max_disks
;
7492 struct md_rdev
*rdev
;
7493 struct disk_info
*disk
;
7497 struct r5worker_group
*new_group
;
7500 if (mddev
->new_level
!= 5
7501 && mddev
->new_level
!= 4
7502 && mddev
->new_level
!= 6) {
7503 pr_warn("md/raid:%s: raid level not set to 4/5/6 (%d)\n",
7504 mdname(mddev
), mddev
->new_level
);
7505 return ERR_PTR(-EIO
);
7507 if ((mddev
->new_level
== 5
7508 && !algorithm_valid_raid5(mddev
->new_layout
)) ||
7509 (mddev
->new_level
== 6
7510 && !algorithm_valid_raid6(mddev
->new_layout
))) {
7511 pr_warn("md/raid:%s: layout %d not supported\n",
7512 mdname(mddev
), mddev
->new_layout
);
7513 return ERR_PTR(-EIO
);
7515 if (mddev
->new_level
== 6 && mddev
->raid_disks
< 4) {
7516 pr_warn("md/raid:%s: not enough configured devices (%d, minimum 4)\n",
7517 mdname(mddev
), mddev
->raid_disks
);
7518 return ERR_PTR(-EINVAL
);
7521 if (!mddev
->new_chunk_sectors
||
7522 (mddev
->new_chunk_sectors
<< 9) % PAGE_SIZE
||
7523 !is_power_of_2(mddev
->new_chunk_sectors
)) {
7524 pr_warn("md/raid:%s: invalid chunk size %d\n",
7525 mdname(mddev
), mddev
->new_chunk_sectors
<< 9);
7526 return ERR_PTR(-EINVAL
);
7529 conf
= kzalloc(sizeof(struct r5conf
), GFP_KERNEL
);
7533 #if PAGE_SIZE != DEFAULT_STRIPE_SIZE
7534 conf
->stripe_size
= DEFAULT_STRIPE_SIZE
;
7535 conf
->stripe_shift
= ilog2(DEFAULT_STRIPE_SIZE
) - 9;
7536 conf
->stripe_sectors
= DEFAULT_STRIPE_SIZE
>> 9;
7538 INIT_LIST_HEAD(&conf
->free_list
);
7539 INIT_LIST_HEAD(&conf
->pending_list
);
7540 conf
->pending_data
= kcalloc(PENDING_IO_MAX
,
7541 sizeof(struct r5pending_data
),
7543 if (!conf
->pending_data
)
7545 for (i
= 0; i
< PENDING_IO_MAX
; i
++)
7546 list_add(&conf
->pending_data
[i
].sibling
, &conf
->free_list
);
7547 /* Don't enable multi-threading by default*/
7548 if (!alloc_thread_groups(conf
, 0, &group_cnt
, &new_group
)) {
7549 conf
->group_cnt
= group_cnt
;
7550 conf
->worker_cnt_per_group
= 0;
7551 conf
->worker_groups
= new_group
;
7554 spin_lock_init(&conf
->device_lock
);
7555 seqcount_spinlock_init(&conf
->gen_lock
, &conf
->device_lock
);
7556 mutex_init(&conf
->cache_size_mutex
);
7558 init_waitqueue_head(&conf
->wait_for_quiescent
);
7559 init_waitqueue_head(&conf
->wait_for_stripe
);
7560 init_waitqueue_head(&conf
->wait_for_overlap
);
7561 INIT_LIST_HEAD(&conf
->handle_list
);
7562 INIT_LIST_HEAD(&conf
->loprio_list
);
7563 INIT_LIST_HEAD(&conf
->hold_list
);
7564 INIT_LIST_HEAD(&conf
->delayed_list
);
7565 INIT_LIST_HEAD(&conf
->bitmap_list
);
7566 init_llist_head(&conf
->released_stripes
);
7567 atomic_set(&conf
->active_stripes
, 0);
7568 atomic_set(&conf
->preread_active_stripes
, 0);
7569 atomic_set(&conf
->active_aligned_reads
, 0);
7570 spin_lock_init(&conf
->pending_bios_lock
);
7571 conf
->batch_bio_dispatch
= true;
7572 rdev_for_each(rdev
, mddev
) {
7573 if (test_bit(Journal
, &rdev
->flags
))
7575 if (bdev_nonrot(rdev
->bdev
)) {
7576 conf
->batch_bio_dispatch
= false;
7581 conf
->bypass_threshold
= BYPASS_THRESHOLD
;
7582 conf
->recovery_disabled
= mddev
->recovery_disabled
- 1;
7584 conf
->raid_disks
= mddev
->raid_disks
;
7585 if (mddev
->reshape_position
== MaxSector
)
7586 conf
->previous_raid_disks
= mddev
->raid_disks
;
7588 conf
->previous_raid_disks
= mddev
->raid_disks
- mddev
->delta_disks
;
7589 max_disks
= max(conf
->raid_disks
, conf
->previous_raid_disks
);
7591 conf
->disks
= kcalloc(max_disks
, sizeof(struct disk_info
),
7597 for (i
= 0; i
< max_disks
; i
++) {
7598 conf
->disks
[i
].extra_page
= alloc_page(GFP_KERNEL
);
7599 if (!conf
->disks
[i
].extra_page
)
7603 ret
= bioset_init(&conf
->bio_split
, BIO_POOL_SIZE
, 0, 0);
7606 conf
->mddev
= mddev
;
7609 conf
->stripe_hashtbl
= kzalloc(PAGE_SIZE
, GFP_KERNEL
);
7610 if (!conf
->stripe_hashtbl
)
7613 /* We init hash_locks[0] separately to that it can be used
7614 * as the reference lock in the spin_lock_nest_lock() call
7615 * in lock_all_device_hash_locks_irq in order to convince
7616 * lockdep that we know what we are doing.
7618 spin_lock_init(conf
->hash_locks
);
7619 for (i
= 1; i
< NR_STRIPE_HASH_LOCKS
; i
++)
7620 spin_lock_init(conf
->hash_locks
+ i
);
7622 for (i
= 0; i
< NR_STRIPE_HASH_LOCKS
; i
++)
7623 INIT_LIST_HEAD(conf
->inactive_list
+ i
);
7625 for (i
= 0; i
< NR_STRIPE_HASH_LOCKS
; i
++)
7626 INIT_LIST_HEAD(conf
->temp_inactive_list
+ i
);
7628 atomic_set(&conf
->r5c_cached_full_stripes
, 0);
7629 INIT_LIST_HEAD(&conf
->r5c_full_stripe_list
);
7630 atomic_set(&conf
->r5c_cached_partial_stripes
, 0);
7631 INIT_LIST_HEAD(&conf
->r5c_partial_stripe_list
);
7632 atomic_set(&conf
->r5c_flushing_full_stripes
, 0);
7633 atomic_set(&conf
->r5c_flushing_partial_stripes
, 0);
7635 conf
->level
= mddev
->new_level
;
7636 conf
->chunk_sectors
= mddev
->new_chunk_sectors
;
7637 ret
= raid5_alloc_percpu(conf
);
7641 pr_debug("raid456: run(%s) called.\n", mdname(mddev
));
7644 rdev_for_each(rdev
, mddev
) {
7645 raid_disk
= rdev
->raid_disk
;
7646 if (raid_disk
>= max_disks
7647 || raid_disk
< 0 || test_bit(Journal
, &rdev
->flags
))
7649 disk
= conf
->disks
+ raid_disk
;
7651 if (test_bit(Replacement
, &rdev
->flags
)) {
7652 if (disk
->replacement
)
7654 RCU_INIT_POINTER(disk
->replacement
, rdev
);
7658 RCU_INIT_POINTER(disk
->rdev
, rdev
);
7661 if (test_bit(In_sync
, &rdev
->flags
)) {
7662 pr_info("md/raid:%s: device %pg operational as raid disk %d\n",
7663 mdname(mddev
), rdev
->bdev
, raid_disk
);
7664 } else if (rdev
->saved_raid_disk
!= raid_disk
)
7665 /* Cannot rely on bitmap to complete recovery */
7669 conf
->level
= mddev
->new_level
;
7670 if (conf
->level
== 6) {
7671 conf
->max_degraded
= 2;
7672 if (raid6_call
.xor_syndrome
)
7673 conf
->rmw_level
= PARITY_ENABLE_RMW
;
7675 conf
->rmw_level
= PARITY_DISABLE_RMW
;
7677 conf
->max_degraded
= 1;
7678 conf
->rmw_level
= PARITY_ENABLE_RMW
;
7680 conf
->algorithm
= mddev
->new_layout
;
7681 conf
->reshape_progress
= mddev
->reshape_position
;
7682 if (conf
->reshape_progress
!= MaxSector
) {
7683 conf
->prev_chunk_sectors
= mddev
->chunk_sectors
;
7684 conf
->prev_algo
= mddev
->layout
;
7686 conf
->prev_chunk_sectors
= conf
->chunk_sectors
;
7687 conf
->prev_algo
= conf
->algorithm
;
7690 conf
->min_nr_stripes
= NR_STRIPES
;
7691 if (mddev
->reshape_position
!= MaxSector
) {
7692 int stripes
= max_t(int,
7693 ((mddev
->chunk_sectors
<< 9) / RAID5_STRIPE_SIZE(conf
)) * 4,
7694 ((mddev
->new_chunk_sectors
<< 9) / RAID5_STRIPE_SIZE(conf
)) * 4);
7695 conf
->min_nr_stripes
= max(NR_STRIPES
, stripes
);
7696 if (conf
->min_nr_stripes
!= NR_STRIPES
)
7697 pr_info("md/raid:%s: force stripe size %d for reshape\n",
7698 mdname(mddev
), conf
->min_nr_stripes
);
7700 memory
= conf
->min_nr_stripes
* (sizeof(struct stripe_head
) +
7701 max_disks
* ((sizeof(struct bio
) + PAGE_SIZE
))) / 1024;
7702 atomic_set(&conf
->empty_inactive_list_nr
, NR_STRIPE_HASH_LOCKS
);
7703 if (grow_stripes(conf
, conf
->min_nr_stripes
)) {
7704 pr_warn("md/raid:%s: couldn't allocate %dkB for buffers\n",
7705 mdname(mddev
), memory
);
7709 pr_debug("md/raid:%s: allocated %dkB\n", mdname(mddev
), memory
);
7711 * Losing a stripe head costs more than the time to refill it,
7712 * it reduces the queue depth and so can hurt throughput.
7713 * So set it rather large, scaled by number of devices.
7715 conf
->shrinker
.seeks
= DEFAULT_SEEKS
* conf
->raid_disks
* 4;
7716 conf
->shrinker
.scan_objects
= raid5_cache_scan
;
7717 conf
->shrinker
.count_objects
= raid5_cache_count
;
7718 conf
->shrinker
.batch
= 128;
7719 conf
->shrinker
.flags
= 0;
7720 ret
= register_shrinker(&conf
->shrinker
, "md-raid5:%s", mdname(mddev
));
7722 pr_warn("md/raid:%s: couldn't register shrinker.\n",
7727 sprintf(pers_name
, "raid%d", mddev
->new_level
);
7728 rcu_assign_pointer(conf
->thread
,
7729 md_register_thread(raid5d
, mddev
, pers_name
));
7730 if (!conf
->thread
) {
7731 pr_warn("md/raid:%s: couldn't allocate thread.\n",
7742 return ERR_PTR(ret
);
7745 static int only_parity(int raid_disk
, int algo
, int raid_disks
, int max_degraded
)
7748 case ALGORITHM_PARITY_0
:
7749 if (raid_disk
< max_degraded
)
7752 case ALGORITHM_PARITY_N
:
7753 if (raid_disk
>= raid_disks
- max_degraded
)
7756 case ALGORITHM_PARITY_0_6
:
7757 if (raid_disk
== 0 ||
7758 raid_disk
== raid_disks
- 1)
7761 case ALGORITHM_LEFT_ASYMMETRIC_6
:
7762 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
7763 case ALGORITHM_LEFT_SYMMETRIC_6
:
7764 case ALGORITHM_RIGHT_SYMMETRIC_6
:
7765 if (raid_disk
== raid_disks
- 1)
7771 static void raid5_set_io_opt(struct r5conf
*conf
)
7773 blk_queue_io_opt(conf
->mddev
->queue
, (conf
->chunk_sectors
<< 9) *
7774 (conf
->raid_disks
- conf
->max_degraded
));
7777 static int raid5_run(struct mddev
*mddev
)
7779 struct r5conf
*conf
;
7780 int dirty_parity_disks
= 0;
7781 struct md_rdev
*rdev
;
7782 struct md_rdev
*journal_dev
= NULL
;
7783 sector_t reshape_offset
= 0;
7785 long long min_offset_diff
= 0;
7788 if (mddev_init_writes_pending(mddev
) < 0)
7791 if (mddev
->recovery_cp
!= MaxSector
)
7792 pr_notice("md/raid:%s: not clean -- starting background reconstruction\n",
7795 rdev_for_each(rdev
, mddev
) {
7798 if (test_bit(Journal
, &rdev
->flags
)) {
7802 if (rdev
->raid_disk
< 0)
7804 diff
= (rdev
->new_data_offset
- rdev
->data_offset
);
7806 min_offset_diff
= diff
;
7808 } else if (mddev
->reshape_backwards
&&
7809 diff
< min_offset_diff
)
7810 min_offset_diff
= diff
;
7811 else if (!mddev
->reshape_backwards
&&
7812 diff
> min_offset_diff
)
7813 min_offset_diff
= diff
;
7816 if ((test_bit(MD_HAS_JOURNAL
, &mddev
->flags
) || journal_dev
) &&
7817 (mddev
->bitmap_info
.offset
|| mddev
->bitmap_info
.file
)) {
7818 pr_notice("md/raid:%s: array cannot have both journal and bitmap\n",
7823 if (mddev
->reshape_position
!= MaxSector
) {
7824 /* Check that we can continue the reshape.
7825 * Difficulties arise if the stripe we would write to
7826 * next is at or after the stripe we would read from next.
7827 * For a reshape that changes the number of devices, this
7828 * is only possible for a very short time, and mdadm makes
7829 * sure that time appears to have past before assembling
7830 * the array. So we fail if that time hasn't passed.
7831 * For a reshape that keeps the number of devices the same
7832 * mdadm must be monitoring the reshape can keeping the
7833 * critical areas read-only and backed up. It will start
7834 * the array in read-only mode, so we check for that.
7836 sector_t here_new
, here_old
;
7838 int max_degraded
= (mddev
->level
== 6 ? 2 : 1);
7843 pr_warn("md/raid:%s: don't support reshape with journal - aborting.\n",
7848 if (mddev
->new_level
!= mddev
->level
) {
7849 pr_warn("md/raid:%s: unsupported reshape required - aborting.\n",
7853 old_disks
= mddev
->raid_disks
- mddev
->delta_disks
;
7854 /* reshape_position must be on a new-stripe boundary, and one
7855 * further up in new geometry must map after here in old
7857 * If the chunk sizes are different, then as we perform reshape
7858 * in units of the largest of the two, reshape_position needs
7859 * be a multiple of the largest chunk size times new data disks.
7861 here_new
= mddev
->reshape_position
;
7862 chunk_sectors
= max(mddev
->chunk_sectors
, mddev
->new_chunk_sectors
);
7863 new_data_disks
= mddev
->raid_disks
- max_degraded
;
7864 if (sector_div(here_new
, chunk_sectors
* new_data_disks
)) {
7865 pr_warn("md/raid:%s: reshape_position not on a stripe boundary\n",
7869 reshape_offset
= here_new
* chunk_sectors
;
7870 /* here_new is the stripe we will write to */
7871 here_old
= mddev
->reshape_position
;
7872 sector_div(here_old
, chunk_sectors
* (old_disks
-max_degraded
));
7873 /* here_old is the first stripe that we might need to read
7875 if (mddev
->delta_disks
== 0) {
7876 /* We cannot be sure it is safe to start an in-place
7877 * reshape. It is only safe if user-space is monitoring
7878 * and taking constant backups.
7879 * mdadm always starts a situation like this in
7880 * readonly mode so it can take control before
7881 * allowing any writes. So just check for that.
7883 if (abs(min_offset_diff
) >= mddev
->chunk_sectors
&&
7884 abs(min_offset_diff
) >= mddev
->new_chunk_sectors
)
7885 /* not really in-place - so OK */;
7886 else if (mddev
->ro
== 0) {
7887 pr_warn("md/raid:%s: in-place reshape must be started in read-only mode - aborting\n",
7891 } else if (mddev
->reshape_backwards
7892 ? (here_new
* chunk_sectors
+ min_offset_diff
<=
7893 here_old
* chunk_sectors
)
7894 : (here_new
* chunk_sectors
>=
7895 here_old
* chunk_sectors
+ (-min_offset_diff
))) {
7896 /* Reading from the same stripe as writing to - bad */
7897 pr_warn("md/raid:%s: reshape_position too early for auto-recovery - aborting.\n",
7901 pr_debug("md/raid:%s: reshape will continue\n", mdname(mddev
));
7902 /* OK, we should be able to continue; */
7904 BUG_ON(mddev
->level
!= mddev
->new_level
);
7905 BUG_ON(mddev
->layout
!= mddev
->new_layout
);
7906 BUG_ON(mddev
->chunk_sectors
!= mddev
->new_chunk_sectors
);
7907 BUG_ON(mddev
->delta_disks
!= 0);
7910 if (test_bit(MD_HAS_JOURNAL
, &mddev
->flags
) &&
7911 test_bit(MD_HAS_PPL
, &mddev
->flags
)) {
7912 pr_warn("md/raid:%s: using journal device and PPL not allowed - disabling PPL\n",
7914 clear_bit(MD_HAS_PPL
, &mddev
->flags
);
7915 clear_bit(MD_HAS_MULTIPLE_PPLS
, &mddev
->flags
);
7918 if (mddev
->private == NULL
)
7919 conf
= setup_conf(mddev
);
7921 conf
= mddev
->private;
7924 return PTR_ERR(conf
);
7926 if (test_bit(MD_HAS_JOURNAL
, &mddev
->flags
)) {
7928 pr_warn("md/raid:%s: journal disk is missing, force array readonly\n",
7931 set_disk_ro(mddev
->gendisk
, 1);
7932 } else if (mddev
->recovery_cp
== MaxSector
)
7933 set_bit(MD_JOURNAL_CLEAN
, &mddev
->flags
);
7936 conf
->min_offset_diff
= min_offset_diff
;
7937 rcu_assign_pointer(mddev
->thread
, conf
->thread
);
7938 rcu_assign_pointer(conf
->thread
, NULL
);
7939 mddev
->private = conf
;
7941 for (i
= 0; i
< conf
->raid_disks
&& conf
->previous_raid_disks
;
7943 rdev
= rdev_mdlock_deref(mddev
, conf
->disks
[i
].rdev
);
7944 if (!rdev
&& conf
->disks
[i
].replacement
) {
7945 /* The replacement is all we have yet */
7946 rdev
= rdev_mdlock_deref(mddev
,
7947 conf
->disks
[i
].replacement
);
7948 conf
->disks
[i
].replacement
= NULL
;
7949 clear_bit(Replacement
, &rdev
->flags
);
7950 rcu_assign_pointer(conf
->disks
[i
].rdev
, rdev
);
7954 if (rcu_access_pointer(conf
->disks
[i
].replacement
) &&
7955 conf
->reshape_progress
!= MaxSector
) {
7956 /* replacements and reshape simply do not mix. */
7957 pr_warn("md: cannot handle concurrent replacement and reshape.\n");
7960 if (test_bit(In_sync
, &rdev
->flags
))
7962 /* This disc is not fully in-sync. However if it
7963 * just stored parity (beyond the recovery_offset),
7964 * when we don't need to be concerned about the
7965 * array being dirty.
7966 * When reshape goes 'backwards', we never have
7967 * partially completed devices, so we only need
7968 * to worry about reshape going forwards.
7970 /* Hack because v0.91 doesn't store recovery_offset properly. */
7971 if (mddev
->major_version
== 0 &&
7972 mddev
->minor_version
> 90)
7973 rdev
->recovery_offset
= reshape_offset
;
7975 if (rdev
->recovery_offset
< reshape_offset
) {
7976 /* We need to check old and new layout */
7977 if (!only_parity(rdev
->raid_disk
,
7980 conf
->max_degraded
))
7983 if (!only_parity(rdev
->raid_disk
,
7985 conf
->previous_raid_disks
,
7986 conf
->max_degraded
))
7988 dirty_parity_disks
++;
7992 * 0 for a fully functional array, 1 or 2 for a degraded array.
7994 mddev
->degraded
= raid5_calc_degraded(conf
);
7996 if (has_failed(conf
)) {
7997 pr_crit("md/raid:%s: not enough operational devices (%d/%d failed)\n",
7998 mdname(mddev
), mddev
->degraded
, conf
->raid_disks
);
8002 /* device size must be a multiple of chunk size */
8003 mddev
->dev_sectors
&= ~((sector_t
)mddev
->chunk_sectors
- 1);
8004 mddev
->resync_max_sectors
= mddev
->dev_sectors
;
8006 if (mddev
->degraded
> dirty_parity_disks
&&
8007 mddev
->recovery_cp
!= MaxSector
) {
8008 if (test_bit(MD_HAS_PPL
, &mddev
->flags
))
8009 pr_crit("md/raid:%s: starting dirty degraded array with PPL.\n",
8011 else if (mddev
->ok_start_degraded
)
8012 pr_crit("md/raid:%s: starting dirty degraded array - data corruption possible.\n",
8015 pr_crit("md/raid:%s: cannot start dirty degraded array.\n",
8021 pr_info("md/raid:%s: raid level %d active with %d out of %d devices, algorithm %d\n",
8022 mdname(mddev
), conf
->level
,
8023 mddev
->raid_disks
-mddev
->degraded
, mddev
->raid_disks
,
8026 print_raid5_conf(conf
);
8028 if (conf
->reshape_progress
!= MaxSector
) {
8029 conf
->reshape_safe
= conf
->reshape_progress
;
8030 atomic_set(&conf
->reshape_stripes
, 0);
8031 clear_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
);
8032 clear_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
);
8033 set_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
);
8034 set_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
);
8035 rcu_assign_pointer(mddev
->sync_thread
,
8036 md_register_thread(md_do_sync
, mddev
, "reshape"));
8037 if (!mddev
->sync_thread
)
8041 /* Ok, everything is just fine now */
8042 if (mddev
->to_remove
== &raid5_attrs_group
)
8043 mddev
->to_remove
= NULL
;
8044 else if (mddev
->kobj
.sd
&&
8045 sysfs_create_group(&mddev
->kobj
, &raid5_attrs_group
))
8046 pr_warn("raid5: failed to create sysfs attributes for %s\n",
8048 md_set_array_sectors(mddev
, raid5_size(mddev
, 0, 0));
8052 /* read-ahead size must cover two whole stripes, which
8053 * is 2 * (datadisks) * chunksize where 'n' is the
8054 * number of raid devices
8056 int data_disks
= conf
->previous_raid_disks
- conf
->max_degraded
;
8057 int stripe
= data_disks
*
8058 ((mddev
->chunk_sectors
<< 9) / PAGE_SIZE
);
8060 chunk_size
= mddev
->chunk_sectors
<< 9;
8061 blk_queue_io_min(mddev
->queue
, chunk_size
);
8062 raid5_set_io_opt(conf
);
8063 mddev
->queue
->limits
.raid_partial_stripes_expensive
= 1;
8065 * We can only discard a whole stripe. It doesn't make sense to
8066 * discard data disk but write parity disk
8068 stripe
= stripe
* PAGE_SIZE
;
8069 stripe
= roundup_pow_of_two(stripe
);
8070 mddev
->queue
->limits
.discard_granularity
= stripe
;
8072 blk_queue_max_write_zeroes_sectors(mddev
->queue
, 0);
8074 rdev_for_each(rdev
, mddev
) {
8075 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
8076 rdev
->data_offset
<< 9);
8077 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
8078 rdev
->new_data_offset
<< 9);
8082 * zeroing is required, otherwise data
8083 * could be lost. Consider a scenario: discard a stripe
8084 * (the stripe could be inconsistent if
8085 * discard_zeroes_data is 0); write one disk of the
8086 * stripe (the stripe could be inconsistent again
8087 * depending on which disks are used to calculate
8088 * parity); the disk is broken; The stripe data of this
8091 * We only allow DISCARD if the sysadmin has confirmed that
8092 * only safe devices are in use by setting a module parameter.
8093 * A better idea might be to turn DISCARD into WRITE_ZEROES
8094 * requests, as that is required to be safe.
8096 if (!devices_handle_discard_safely
||
8097 mddev
->queue
->limits
.max_discard_sectors
< (stripe
>> 9) ||
8098 mddev
->queue
->limits
.discard_granularity
< stripe
)
8099 blk_queue_max_discard_sectors(mddev
->queue
, 0);
8102 * Requests require having a bitmap for each stripe.
8103 * Limit the max sectors based on this.
8105 blk_queue_max_hw_sectors(mddev
->queue
,
8106 RAID5_MAX_REQ_STRIPES
<< RAID5_STRIPE_SHIFT(conf
));
8108 /* No restrictions on the number of segments in the request */
8109 blk_queue_max_segments(mddev
->queue
, USHRT_MAX
);
8112 if (log_init(conf
, journal_dev
, raid5_has_ppl(conf
)))
8117 md_unregister_thread(mddev
, &mddev
->thread
);
8118 print_raid5_conf(conf
);
8120 mddev
->private = NULL
;
8121 pr_warn("md/raid:%s: failed to run raid set.\n", mdname(mddev
));
8125 static void raid5_free(struct mddev
*mddev
, void *priv
)
8127 struct r5conf
*conf
= priv
;
8130 mddev
->to_remove
= &raid5_attrs_group
;
8133 static void raid5_status(struct seq_file
*seq
, struct mddev
*mddev
)
8135 struct r5conf
*conf
= mddev
->private;
8138 seq_printf(seq
, " level %d, %dk chunk, algorithm %d", mddev
->level
,
8139 conf
->chunk_sectors
/ 2, mddev
->layout
);
8140 seq_printf (seq
, " [%d/%d] [", conf
->raid_disks
, conf
->raid_disks
- mddev
->degraded
);
8142 for (i
= 0; i
< conf
->raid_disks
; i
++) {
8143 struct md_rdev
*rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
8144 seq_printf (seq
, "%s", rdev
&& test_bit(In_sync
, &rdev
->flags
) ? "U" : "_");
8147 seq_printf (seq
, "]");
8150 static void print_raid5_conf (struct r5conf
*conf
)
8152 struct md_rdev
*rdev
;
8155 pr_debug("RAID conf printout:\n");
8157 pr_debug("(conf==NULL)\n");
8160 pr_debug(" --- level:%d rd:%d wd:%d\n", conf
->level
,
8162 conf
->raid_disks
- conf
->mddev
->degraded
);
8165 for (i
= 0; i
< conf
->raid_disks
; i
++) {
8166 rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
8168 pr_debug(" disk %d, o:%d, dev:%pg\n",
8169 i
, !test_bit(Faulty
, &rdev
->flags
),
8175 static int raid5_spare_active(struct mddev
*mddev
)
8178 struct r5conf
*conf
= mddev
->private;
8179 struct md_rdev
*rdev
, *replacement
;
8181 unsigned long flags
;
8183 for (i
= 0; i
< conf
->raid_disks
; i
++) {
8184 rdev
= rdev_mdlock_deref(mddev
, conf
->disks
[i
].rdev
);
8185 replacement
= rdev_mdlock_deref(mddev
,
8186 conf
->disks
[i
].replacement
);
8188 && replacement
->recovery_offset
== MaxSector
8189 && !test_bit(Faulty
, &replacement
->flags
)
8190 && !test_and_set_bit(In_sync
, &replacement
->flags
)) {
8191 /* Replacement has just become active. */
8193 || !test_and_clear_bit(In_sync
, &rdev
->flags
))
8196 /* Replaced device not technically faulty,
8197 * but we need to be sure it gets removed
8198 * and never re-added.
8200 set_bit(Faulty
, &rdev
->flags
);
8201 sysfs_notify_dirent_safe(
8204 sysfs_notify_dirent_safe(replacement
->sysfs_state
);
8206 && rdev
->recovery_offset
== MaxSector
8207 && !test_bit(Faulty
, &rdev
->flags
)
8208 && !test_and_set_bit(In_sync
, &rdev
->flags
)) {
8210 sysfs_notify_dirent_safe(rdev
->sysfs_state
);
8213 spin_lock_irqsave(&conf
->device_lock
, flags
);
8214 mddev
->degraded
= raid5_calc_degraded(conf
);
8215 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
8216 print_raid5_conf(conf
);
8220 static int raid5_remove_disk(struct mddev
*mddev
, struct md_rdev
*rdev
)
8222 struct r5conf
*conf
= mddev
->private;
8224 int number
= rdev
->raid_disk
;
8225 struct md_rdev __rcu
**rdevp
;
8226 struct disk_info
*p
;
8227 struct md_rdev
*tmp
;
8229 print_raid5_conf(conf
);
8230 if (test_bit(Journal
, &rdev
->flags
) && conf
->log
) {
8232 * we can't wait pending write here, as this is called in
8233 * raid5d, wait will deadlock.
8234 * neilb: there is no locking about new writes here,
8235 * so this cannot be safe.
8237 if (atomic_read(&conf
->active_stripes
) ||
8238 atomic_read(&conf
->r5c_cached_full_stripes
) ||
8239 atomic_read(&conf
->r5c_cached_partial_stripes
)) {
8245 if (unlikely(number
>= conf
->pool_size
))
8247 p
= conf
->disks
+ number
;
8248 if (rdev
== rcu_access_pointer(p
->rdev
))
8250 else if (rdev
== rcu_access_pointer(p
->replacement
))
8251 rdevp
= &p
->replacement
;
8255 if (number
>= conf
->raid_disks
&&
8256 conf
->reshape_progress
== MaxSector
)
8257 clear_bit(In_sync
, &rdev
->flags
);
8259 if (test_bit(In_sync
, &rdev
->flags
) ||
8260 atomic_read(&rdev
->nr_pending
)) {
8264 /* Only remove non-faulty devices if recovery
8267 if (!test_bit(Faulty
, &rdev
->flags
) &&
8268 mddev
->recovery_disabled
!= conf
->recovery_disabled
&&
8269 !has_failed(conf
) &&
8270 (!rcu_access_pointer(p
->replacement
) ||
8271 rcu_access_pointer(p
->replacement
) == rdev
) &&
8272 number
< conf
->raid_disks
) {
8277 if (!test_bit(RemoveSynchronized
, &rdev
->flags
)) {
8278 lockdep_assert_held(&mddev
->reconfig_mutex
);
8280 if (atomic_read(&rdev
->nr_pending
)) {
8281 /* lost the race, try later */
8283 rcu_assign_pointer(*rdevp
, rdev
);
8287 err
= log_modify(conf
, rdev
, false);
8292 tmp
= rcu_access_pointer(p
->replacement
);
8294 /* We must have just cleared 'rdev' */
8295 rcu_assign_pointer(p
->rdev
, tmp
);
8296 clear_bit(Replacement
, &tmp
->flags
);
8297 smp_mb(); /* Make sure other CPUs may see both as identical
8298 * but will never see neither - if they are careful
8300 rcu_assign_pointer(p
->replacement
, NULL
);
8303 err
= log_modify(conf
, tmp
, true);
8306 clear_bit(WantReplacement
, &rdev
->flags
);
8309 print_raid5_conf(conf
);
8313 static int raid5_add_disk(struct mddev
*mddev
, struct md_rdev
*rdev
)
8315 struct r5conf
*conf
= mddev
->private;
8316 int ret
, err
= -EEXIST
;
8318 struct disk_info
*p
;
8319 struct md_rdev
*tmp
;
8321 int last
= conf
->raid_disks
- 1;
8323 if (test_bit(Journal
, &rdev
->flags
)) {
8327 rdev
->raid_disk
= 0;
8329 * The array is in readonly mode if journal is missing, so no
8330 * write requests running. We should be safe
8332 ret
= log_init(conf
, rdev
, false);
8336 ret
= r5l_start(conf
->log
);
8342 if (mddev
->recovery_disabled
== conf
->recovery_disabled
)
8345 if (rdev
->saved_raid_disk
< 0 && has_failed(conf
))
8346 /* no point adding a device */
8349 if (rdev
->raid_disk
>= 0)
8350 first
= last
= rdev
->raid_disk
;
8353 * find the disk ... but prefer rdev->saved_raid_disk
8356 if (rdev
->saved_raid_disk
>= first
&&
8357 rdev
->saved_raid_disk
<= last
&&
8358 conf
->disks
[rdev
->saved_raid_disk
].rdev
== NULL
)
8359 first
= rdev
->saved_raid_disk
;
8361 for (disk
= first
; disk
<= last
; disk
++) {
8362 p
= conf
->disks
+ disk
;
8363 if (p
->rdev
== NULL
) {
8364 clear_bit(In_sync
, &rdev
->flags
);
8365 rdev
->raid_disk
= disk
;
8366 if (rdev
->saved_raid_disk
!= disk
)
8368 rcu_assign_pointer(p
->rdev
, rdev
);
8370 err
= log_modify(conf
, rdev
, true);
8375 for (disk
= first
; disk
<= last
; disk
++) {
8376 p
= conf
->disks
+ disk
;
8377 tmp
= rdev_mdlock_deref(mddev
, p
->rdev
);
8378 if (test_bit(WantReplacement
, &tmp
->flags
) &&
8379 mddev
->reshape_position
== MaxSector
&&
8380 p
->replacement
== NULL
) {
8381 clear_bit(In_sync
, &rdev
->flags
);
8382 set_bit(Replacement
, &rdev
->flags
);
8383 rdev
->raid_disk
= disk
;
8386 rcu_assign_pointer(p
->replacement
, rdev
);
8391 print_raid5_conf(conf
);
8395 static int raid5_resize(struct mddev
*mddev
, sector_t sectors
)
8397 /* no resync is happening, and there is enough space
8398 * on all devices, so we can resize.
8399 * We need to make sure resync covers any new space.
8400 * If the array is shrinking we should possibly wait until
8401 * any io in the removed space completes, but it hardly seems
8405 struct r5conf
*conf
= mddev
->private;
8407 if (raid5_has_log(conf
) || raid5_has_ppl(conf
))
8409 sectors
&= ~((sector_t
)conf
->chunk_sectors
- 1);
8410 newsize
= raid5_size(mddev
, sectors
, mddev
->raid_disks
);
8411 if (mddev
->external_size
&&
8412 mddev
->array_sectors
> newsize
)
8414 if (mddev
->bitmap
) {
8415 int ret
= md_bitmap_resize(mddev
->bitmap
, sectors
, 0, 0);
8419 md_set_array_sectors(mddev
, newsize
);
8420 if (sectors
> mddev
->dev_sectors
&&
8421 mddev
->recovery_cp
> mddev
->dev_sectors
) {
8422 mddev
->recovery_cp
= mddev
->dev_sectors
;
8423 set_bit(MD_RECOVERY_NEEDED
, &mddev
->recovery
);
8425 mddev
->dev_sectors
= sectors
;
8426 mddev
->resync_max_sectors
= sectors
;
8430 static int check_stripe_cache(struct mddev
*mddev
)
8432 /* Can only proceed if there are plenty of stripe_heads.
8433 * We need a minimum of one full stripe,, and for sensible progress
8434 * it is best to have about 4 times that.
8435 * If we require 4 times, then the default 256 4K stripe_heads will
8436 * allow for chunk sizes up to 256K, which is probably OK.
8437 * If the chunk size is greater, user-space should request more
8438 * stripe_heads first.
8440 struct r5conf
*conf
= mddev
->private;
8441 if (((mddev
->chunk_sectors
<< 9) / RAID5_STRIPE_SIZE(conf
)) * 4
8442 > conf
->min_nr_stripes
||
8443 ((mddev
->new_chunk_sectors
<< 9) / RAID5_STRIPE_SIZE(conf
)) * 4
8444 > conf
->min_nr_stripes
) {
8445 pr_warn("md/raid:%s: reshape: not enough stripes. Needed %lu\n",
8447 ((max(mddev
->chunk_sectors
, mddev
->new_chunk_sectors
) << 9)
8448 / RAID5_STRIPE_SIZE(conf
))*4);
8454 static int check_reshape(struct mddev
*mddev
)
8456 struct r5conf
*conf
= mddev
->private;
8458 if (raid5_has_log(conf
) || raid5_has_ppl(conf
))
8460 if (mddev
->delta_disks
== 0 &&
8461 mddev
->new_layout
== mddev
->layout
&&
8462 mddev
->new_chunk_sectors
== mddev
->chunk_sectors
)
8463 return 0; /* nothing to do */
8464 if (has_failed(conf
))
8466 if (mddev
->delta_disks
< 0 && mddev
->reshape_position
== MaxSector
) {
8467 /* We might be able to shrink, but the devices must
8468 * be made bigger first.
8469 * For raid6, 4 is the minimum size.
8470 * Otherwise 2 is the minimum
8473 if (mddev
->level
== 6)
8475 if (mddev
->raid_disks
+ mddev
->delta_disks
< min
)
8479 if (!check_stripe_cache(mddev
))
8482 if (mddev
->new_chunk_sectors
> mddev
->chunk_sectors
||
8483 mddev
->delta_disks
> 0)
8484 if (resize_chunks(conf
,
8485 conf
->previous_raid_disks
8486 + max(0, mddev
->delta_disks
),
8487 max(mddev
->new_chunk_sectors
,
8488 mddev
->chunk_sectors
)
8492 if (conf
->previous_raid_disks
+ mddev
->delta_disks
<= conf
->pool_size
)
8493 return 0; /* never bother to shrink */
8494 return resize_stripes(conf
, (conf
->previous_raid_disks
8495 + mddev
->delta_disks
));
8498 static int raid5_start_reshape(struct mddev
*mddev
)
8500 struct r5conf
*conf
= mddev
->private;
8501 struct md_rdev
*rdev
;
8504 unsigned long flags
;
8506 if (test_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
))
8509 if (!check_stripe_cache(mddev
))
8512 if (has_failed(conf
))
8515 /* raid5 can't handle concurrent reshape and recovery */
8516 if (mddev
->recovery_cp
< MaxSector
)
8518 for (i
= 0; i
< conf
->raid_disks
; i
++)
8519 if (rdev_mdlock_deref(mddev
, conf
->disks
[i
].replacement
))
8522 rdev_for_each(rdev
, mddev
) {
8523 if (!test_bit(In_sync
, &rdev
->flags
)
8524 && !test_bit(Faulty
, &rdev
->flags
))
8528 if (spares
- mddev
->degraded
< mddev
->delta_disks
- conf
->max_degraded
)
8529 /* Not enough devices even to make a degraded array
8534 /* Refuse to reduce size of the array. Any reductions in
8535 * array size must be through explicit setting of array_size
8538 if (raid5_size(mddev
, 0, conf
->raid_disks
+ mddev
->delta_disks
)
8539 < mddev
->array_sectors
) {
8540 pr_warn("md/raid:%s: array size must be reduced before number of disks\n",
8545 atomic_set(&conf
->reshape_stripes
, 0);
8546 spin_lock_irq(&conf
->device_lock
);
8547 write_seqcount_begin(&conf
->gen_lock
);
8548 conf
->previous_raid_disks
= conf
->raid_disks
;
8549 conf
->raid_disks
+= mddev
->delta_disks
;
8550 conf
->prev_chunk_sectors
= conf
->chunk_sectors
;
8551 conf
->chunk_sectors
= mddev
->new_chunk_sectors
;
8552 conf
->prev_algo
= conf
->algorithm
;
8553 conf
->algorithm
= mddev
->new_layout
;
8555 /* Code that selects data_offset needs to see the generation update
8556 * if reshape_progress has been set - so a memory barrier needed.
8559 if (mddev
->reshape_backwards
)
8560 conf
->reshape_progress
= raid5_size(mddev
, 0, 0);
8562 conf
->reshape_progress
= 0;
8563 conf
->reshape_safe
= conf
->reshape_progress
;
8564 write_seqcount_end(&conf
->gen_lock
);
8565 spin_unlock_irq(&conf
->device_lock
);
8567 /* Now make sure any requests that proceeded on the assumption
8568 * the reshape wasn't running - like Discard or Read - have
8571 mddev_suspend(mddev
);
8572 mddev_resume(mddev
);
8574 /* Add some new drives, as many as will fit.
8575 * We know there are enough to make the newly sized array work.
8576 * Don't add devices if we are reducing the number of
8577 * devices in the array. This is because it is not possible
8578 * to correctly record the "partially reconstructed" state of
8579 * such devices during the reshape and confusion could result.
8581 if (mddev
->delta_disks
>= 0) {
8582 rdev_for_each(rdev
, mddev
)
8583 if (rdev
->raid_disk
< 0 &&
8584 !test_bit(Faulty
, &rdev
->flags
)) {
8585 if (raid5_add_disk(mddev
, rdev
) == 0) {
8587 >= conf
->previous_raid_disks
)
8588 set_bit(In_sync
, &rdev
->flags
);
8590 rdev
->recovery_offset
= 0;
8592 /* Failure here is OK */
8593 sysfs_link_rdev(mddev
, rdev
);
8595 } else if (rdev
->raid_disk
>= conf
->previous_raid_disks
8596 && !test_bit(Faulty
, &rdev
->flags
)) {
8597 /* This is a spare that was manually added */
8598 set_bit(In_sync
, &rdev
->flags
);
8601 /* When a reshape changes the number of devices,
8602 * ->degraded is measured against the larger of the
8603 * pre and post number of devices.
8605 spin_lock_irqsave(&conf
->device_lock
, flags
);
8606 mddev
->degraded
= raid5_calc_degraded(conf
);
8607 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
8609 mddev
->raid_disks
= conf
->raid_disks
;
8610 mddev
->reshape_position
= conf
->reshape_progress
;
8611 set_bit(MD_SB_CHANGE_DEVS
, &mddev
->sb_flags
);
8613 clear_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
);
8614 clear_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
);
8615 clear_bit(MD_RECOVERY_DONE
, &mddev
->recovery
);
8616 set_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
);
8617 set_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
);
8618 rcu_assign_pointer(mddev
->sync_thread
,
8619 md_register_thread(md_do_sync
, mddev
, "reshape"));
8620 if (!mddev
->sync_thread
) {
8621 mddev
->recovery
= 0;
8622 spin_lock_irq(&conf
->device_lock
);
8623 write_seqcount_begin(&conf
->gen_lock
);
8624 mddev
->raid_disks
= conf
->raid_disks
= conf
->previous_raid_disks
;
8625 mddev
->new_chunk_sectors
=
8626 conf
->chunk_sectors
= conf
->prev_chunk_sectors
;
8627 mddev
->new_layout
= conf
->algorithm
= conf
->prev_algo
;
8628 rdev_for_each(rdev
, mddev
)
8629 rdev
->new_data_offset
= rdev
->data_offset
;
8631 conf
->generation
--;
8632 conf
->reshape_progress
= MaxSector
;
8633 mddev
->reshape_position
= MaxSector
;
8634 write_seqcount_end(&conf
->gen_lock
);
8635 spin_unlock_irq(&conf
->device_lock
);
8638 conf
->reshape_checkpoint
= jiffies
;
8639 md_wakeup_thread(mddev
->sync_thread
);
8644 /* This is called from the reshape thread and should make any
8645 * changes needed in 'conf'
8647 static void end_reshape(struct r5conf
*conf
)
8650 if (!test_bit(MD_RECOVERY_INTR
, &conf
->mddev
->recovery
)) {
8651 struct md_rdev
*rdev
;
8653 spin_lock_irq(&conf
->device_lock
);
8654 conf
->previous_raid_disks
= conf
->raid_disks
;
8655 md_finish_reshape(conf
->mddev
);
8657 conf
->reshape_progress
= MaxSector
;
8658 conf
->mddev
->reshape_position
= MaxSector
;
8659 rdev_for_each(rdev
, conf
->mddev
)
8660 if (rdev
->raid_disk
>= 0 &&
8661 !test_bit(Journal
, &rdev
->flags
) &&
8662 !test_bit(In_sync
, &rdev
->flags
))
8663 rdev
->recovery_offset
= MaxSector
;
8664 spin_unlock_irq(&conf
->device_lock
);
8665 wake_up(&conf
->wait_for_overlap
);
8667 if (conf
->mddev
->queue
)
8668 raid5_set_io_opt(conf
);
8672 /* This is called from the raid5d thread with mddev_lock held.
8673 * It makes config changes to the device.
8675 static void raid5_finish_reshape(struct mddev
*mddev
)
8677 struct r5conf
*conf
= mddev
->private;
8678 struct md_rdev
*rdev
;
8680 if (!test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
)) {
8682 if (mddev
->delta_disks
<= 0) {
8684 spin_lock_irq(&conf
->device_lock
);
8685 mddev
->degraded
= raid5_calc_degraded(conf
);
8686 spin_unlock_irq(&conf
->device_lock
);
8687 for (d
= conf
->raid_disks
;
8688 d
< conf
->raid_disks
- mddev
->delta_disks
;
8690 rdev
= rdev_mdlock_deref(mddev
,
8691 conf
->disks
[d
].rdev
);
8693 clear_bit(In_sync
, &rdev
->flags
);
8694 rdev
= rdev_mdlock_deref(mddev
,
8695 conf
->disks
[d
].replacement
);
8697 clear_bit(In_sync
, &rdev
->flags
);
8700 mddev
->layout
= conf
->algorithm
;
8701 mddev
->chunk_sectors
= conf
->chunk_sectors
;
8702 mddev
->reshape_position
= MaxSector
;
8703 mddev
->delta_disks
= 0;
8704 mddev
->reshape_backwards
= 0;
8708 static void raid5_quiesce(struct mddev
*mddev
, int quiesce
)
8710 struct r5conf
*conf
= mddev
->private;
8713 /* stop all writes */
8714 lock_all_device_hash_locks_irq(conf
);
8715 /* '2' tells resync/reshape to pause so that all
8716 * active stripes can drain
8718 r5c_flush_cache(conf
, INT_MAX
);
8719 /* need a memory barrier to make sure read_one_chunk() sees
8720 * quiesce started and reverts to slow (locked) path.
8722 smp_store_release(&conf
->quiesce
, 2);
8723 wait_event_cmd(conf
->wait_for_quiescent
,
8724 atomic_read(&conf
->active_stripes
) == 0 &&
8725 atomic_read(&conf
->active_aligned_reads
) == 0,
8726 unlock_all_device_hash_locks_irq(conf
),
8727 lock_all_device_hash_locks_irq(conf
));
8729 unlock_all_device_hash_locks_irq(conf
);
8730 /* allow reshape to continue */
8731 wake_up(&conf
->wait_for_overlap
);
8733 /* re-enable writes */
8734 lock_all_device_hash_locks_irq(conf
);
8736 wake_up(&conf
->wait_for_quiescent
);
8737 wake_up(&conf
->wait_for_overlap
);
8738 unlock_all_device_hash_locks_irq(conf
);
8740 log_quiesce(conf
, quiesce
);
8743 static void *raid45_takeover_raid0(struct mddev
*mddev
, int level
)
8745 struct r0conf
*raid0_conf
= mddev
->private;
8748 /* for raid0 takeover only one zone is supported */
8749 if (raid0_conf
->nr_strip_zones
> 1) {
8750 pr_warn("md/raid:%s: cannot takeover raid0 with more than one zone.\n",
8752 return ERR_PTR(-EINVAL
);
8755 sectors
= raid0_conf
->strip_zone
[0].zone_end
;
8756 sector_div(sectors
, raid0_conf
->strip_zone
[0].nb_dev
);
8757 mddev
->dev_sectors
= sectors
;
8758 mddev
->new_level
= level
;
8759 mddev
->new_layout
= ALGORITHM_PARITY_N
;
8760 mddev
->new_chunk_sectors
= mddev
->chunk_sectors
;
8761 mddev
->raid_disks
+= 1;
8762 mddev
->delta_disks
= 1;
8763 /* make sure it will be not marked as dirty */
8764 mddev
->recovery_cp
= MaxSector
;
8766 return setup_conf(mddev
);
8769 static void *raid5_takeover_raid1(struct mddev
*mddev
)
8774 if (mddev
->raid_disks
!= 2 ||
8775 mddev
->degraded
> 1)
8776 return ERR_PTR(-EINVAL
);
8778 /* Should check if there are write-behind devices? */
8780 chunksect
= 64*2; /* 64K by default */
8782 /* The array must be an exact multiple of chunksize */
8783 while (chunksect
&& (mddev
->array_sectors
& (chunksect
-1)))
8786 if ((chunksect
<<9) < RAID5_STRIPE_SIZE((struct r5conf
*)mddev
->private))
8787 /* array size does not allow a suitable chunk size */
8788 return ERR_PTR(-EINVAL
);
8790 mddev
->new_level
= 5;
8791 mddev
->new_layout
= ALGORITHM_LEFT_SYMMETRIC
;
8792 mddev
->new_chunk_sectors
= chunksect
;
8794 ret
= setup_conf(mddev
);
8796 mddev_clear_unsupported_flags(mddev
,
8797 UNSUPPORTED_MDDEV_FLAGS
);
8801 static void *raid5_takeover_raid6(struct mddev
*mddev
)
8805 switch (mddev
->layout
) {
8806 case ALGORITHM_LEFT_ASYMMETRIC_6
:
8807 new_layout
= ALGORITHM_LEFT_ASYMMETRIC
;
8809 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
8810 new_layout
= ALGORITHM_RIGHT_ASYMMETRIC
;
8812 case ALGORITHM_LEFT_SYMMETRIC_6
:
8813 new_layout
= ALGORITHM_LEFT_SYMMETRIC
;
8815 case ALGORITHM_RIGHT_SYMMETRIC_6
:
8816 new_layout
= ALGORITHM_RIGHT_SYMMETRIC
;
8818 case ALGORITHM_PARITY_0_6
:
8819 new_layout
= ALGORITHM_PARITY_0
;
8821 case ALGORITHM_PARITY_N
:
8822 new_layout
= ALGORITHM_PARITY_N
;
8825 return ERR_PTR(-EINVAL
);
8827 mddev
->new_level
= 5;
8828 mddev
->new_layout
= new_layout
;
8829 mddev
->delta_disks
= -1;
8830 mddev
->raid_disks
-= 1;
8831 return setup_conf(mddev
);
8834 static int raid5_check_reshape(struct mddev
*mddev
)
8836 /* For a 2-drive array, the layout and chunk size can be changed
8837 * immediately as not restriping is needed.
8838 * For larger arrays we record the new value - after validation
8839 * to be used by a reshape pass.
8841 struct r5conf
*conf
= mddev
->private;
8842 int new_chunk
= mddev
->new_chunk_sectors
;
8844 if (mddev
->new_layout
>= 0 && !algorithm_valid_raid5(mddev
->new_layout
))
8846 if (new_chunk
> 0) {
8847 if (!is_power_of_2(new_chunk
))
8849 if (new_chunk
< (PAGE_SIZE
>>9))
8851 if (mddev
->array_sectors
& (new_chunk
-1))
8852 /* not factor of array size */
8856 /* They look valid */
8858 if (mddev
->raid_disks
== 2) {
8859 /* can make the change immediately */
8860 if (mddev
->new_layout
>= 0) {
8861 conf
->algorithm
= mddev
->new_layout
;
8862 mddev
->layout
= mddev
->new_layout
;
8864 if (new_chunk
> 0) {
8865 conf
->chunk_sectors
= new_chunk
;
8866 mddev
->chunk_sectors
= new_chunk
;
8868 set_bit(MD_SB_CHANGE_DEVS
, &mddev
->sb_flags
);
8869 md_wakeup_thread(mddev
->thread
);
8871 return check_reshape(mddev
);
8874 static int raid6_check_reshape(struct mddev
*mddev
)
8876 int new_chunk
= mddev
->new_chunk_sectors
;
8878 if (mddev
->new_layout
>= 0 && !algorithm_valid_raid6(mddev
->new_layout
))
8880 if (new_chunk
> 0) {
8881 if (!is_power_of_2(new_chunk
))
8883 if (new_chunk
< (PAGE_SIZE
>> 9))
8885 if (mddev
->array_sectors
& (new_chunk
-1))
8886 /* not factor of array size */
8890 /* They look valid */
8891 return check_reshape(mddev
);
8894 static void *raid5_takeover(struct mddev
*mddev
)
8896 /* raid5 can take over:
8897 * raid0 - if there is only one strip zone - make it a raid4 layout
8898 * raid1 - if there are two drives. We need to know the chunk size
8899 * raid4 - trivial - just use a raid4 layout.
8900 * raid6 - Providing it is a *_6 layout
8902 if (mddev
->level
== 0)
8903 return raid45_takeover_raid0(mddev
, 5);
8904 if (mddev
->level
== 1)
8905 return raid5_takeover_raid1(mddev
);
8906 if (mddev
->level
== 4) {
8907 mddev
->new_layout
= ALGORITHM_PARITY_N
;
8908 mddev
->new_level
= 5;
8909 return setup_conf(mddev
);
8911 if (mddev
->level
== 6)
8912 return raid5_takeover_raid6(mddev
);
8914 return ERR_PTR(-EINVAL
);
8917 static void *raid4_takeover(struct mddev
*mddev
)
8919 /* raid4 can take over:
8920 * raid0 - if there is only one strip zone
8921 * raid5 - if layout is right
8923 if (mddev
->level
== 0)
8924 return raid45_takeover_raid0(mddev
, 4);
8925 if (mddev
->level
== 5 &&
8926 mddev
->layout
== ALGORITHM_PARITY_N
) {
8927 mddev
->new_layout
= 0;
8928 mddev
->new_level
= 4;
8929 return setup_conf(mddev
);
8931 return ERR_PTR(-EINVAL
);
8934 static struct md_personality raid5_personality
;
8936 static void *raid6_takeover(struct mddev
*mddev
)
8938 /* Currently can only take over a raid5. We map the
8939 * personality to an equivalent raid6 personality
8940 * with the Q block at the end.
8944 if (mddev
->pers
!= &raid5_personality
)
8945 return ERR_PTR(-EINVAL
);
8946 if (mddev
->degraded
> 1)
8947 return ERR_PTR(-EINVAL
);
8948 if (mddev
->raid_disks
> 253)
8949 return ERR_PTR(-EINVAL
);
8950 if (mddev
->raid_disks
< 3)
8951 return ERR_PTR(-EINVAL
);
8953 switch (mddev
->layout
) {
8954 case ALGORITHM_LEFT_ASYMMETRIC
:
8955 new_layout
= ALGORITHM_LEFT_ASYMMETRIC_6
;
8957 case ALGORITHM_RIGHT_ASYMMETRIC
:
8958 new_layout
= ALGORITHM_RIGHT_ASYMMETRIC_6
;
8960 case ALGORITHM_LEFT_SYMMETRIC
:
8961 new_layout
= ALGORITHM_LEFT_SYMMETRIC_6
;
8963 case ALGORITHM_RIGHT_SYMMETRIC
:
8964 new_layout
= ALGORITHM_RIGHT_SYMMETRIC_6
;
8966 case ALGORITHM_PARITY_0
:
8967 new_layout
= ALGORITHM_PARITY_0_6
;
8969 case ALGORITHM_PARITY_N
:
8970 new_layout
= ALGORITHM_PARITY_N
;
8973 return ERR_PTR(-EINVAL
);
8975 mddev
->new_level
= 6;
8976 mddev
->new_layout
= new_layout
;
8977 mddev
->delta_disks
= 1;
8978 mddev
->raid_disks
+= 1;
8979 return setup_conf(mddev
);
8982 static int raid5_change_consistency_policy(struct mddev
*mddev
, const char *buf
)
8984 struct r5conf
*conf
;
8987 err
= mddev_lock(mddev
);
8990 conf
= mddev
->private;
8992 mddev_unlock(mddev
);
8996 if (strncmp(buf
, "ppl", 3) == 0) {
8997 /* ppl only works with RAID 5 */
8998 if (!raid5_has_ppl(conf
) && conf
->level
== 5) {
8999 err
= log_init(conf
, NULL
, true);
9001 err
= resize_stripes(conf
, conf
->pool_size
);
9003 mddev_suspend(mddev
);
9005 mddev_resume(mddev
);
9010 } else if (strncmp(buf
, "resync", 6) == 0) {
9011 if (raid5_has_ppl(conf
)) {
9012 mddev_suspend(mddev
);
9014 mddev_resume(mddev
);
9015 err
= resize_stripes(conf
, conf
->pool_size
);
9016 } else if (test_bit(MD_HAS_JOURNAL
, &conf
->mddev
->flags
) &&
9017 r5l_log_disk_error(conf
)) {
9018 bool journal_dev_exists
= false;
9019 struct md_rdev
*rdev
;
9021 rdev_for_each(rdev
, mddev
)
9022 if (test_bit(Journal
, &rdev
->flags
)) {
9023 journal_dev_exists
= true;
9027 if (!journal_dev_exists
) {
9028 mddev_suspend(mddev
);
9029 clear_bit(MD_HAS_JOURNAL
, &mddev
->flags
);
9030 mddev_resume(mddev
);
9031 } else /* need remove journal device first */
9040 md_update_sb(mddev
, 1);
9042 mddev_unlock(mddev
);
9047 static int raid5_start(struct mddev
*mddev
)
9049 struct r5conf
*conf
= mddev
->private;
9051 return r5l_start(conf
->log
);
9054 static void raid5_prepare_suspend(struct mddev
*mddev
)
9056 struct r5conf
*conf
= mddev
->private;
9058 wait_event(mddev
->sb_wait
, !reshape_inprogress(mddev
) ||
9059 percpu_ref_is_zero(&mddev
->active_io
));
9060 if (percpu_ref_is_zero(&mddev
->active_io
))
9064 * Reshape is not in progress, and array is suspended, io that is
9065 * waiting for reshpape can never be done.
9067 wake_up(&conf
->wait_for_overlap
);
9070 static struct md_personality raid6_personality
=
9074 .owner
= THIS_MODULE
,
9075 .make_request
= raid5_make_request
,
9077 .start
= raid5_start
,
9079 .status
= raid5_status
,
9080 .error_handler
= raid5_error
,
9081 .hot_add_disk
= raid5_add_disk
,
9082 .hot_remove_disk
= raid5_remove_disk
,
9083 .spare_active
= raid5_spare_active
,
9084 .sync_request
= raid5_sync_request
,
9085 .resize
= raid5_resize
,
9087 .check_reshape
= raid6_check_reshape
,
9088 .start_reshape
= raid5_start_reshape
,
9089 .finish_reshape
= raid5_finish_reshape
,
9090 .prepare_suspend
= raid5_prepare_suspend
,
9091 .quiesce
= raid5_quiesce
,
9092 .takeover
= raid6_takeover
,
9093 .change_consistency_policy
= raid5_change_consistency_policy
,
9095 static struct md_personality raid5_personality
=
9099 .owner
= THIS_MODULE
,
9100 .make_request
= raid5_make_request
,
9102 .start
= raid5_start
,
9104 .status
= raid5_status
,
9105 .error_handler
= raid5_error
,
9106 .hot_add_disk
= raid5_add_disk
,
9107 .hot_remove_disk
= raid5_remove_disk
,
9108 .spare_active
= raid5_spare_active
,
9109 .sync_request
= raid5_sync_request
,
9110 .resize
= raid5_resize
,
9112 .check_reshape
= raid5_check_reshape
,
9113 .start_reshape
= raid5_start_reshape
,
9114 .finish_reshape
= raid5_finish_reshape
,
9115 .prepare_suspend
= raid5_prepare_suspend
,
9116 .quiesce
= raid5_quiesce
,
9117 .takeover
= raid5_takeover
,
9118 .change_consistency_policy
= raid5_change_consistency_policy
,
9121 static struct md_personality raid4_personality
=
9125 .owner
= THIS_MODULE
,
9126 .make_request
= raid5_make_request
,
9128 .start
= raid5_start
,
9130 .status
= raid5_status
,
9131 .error_handler
= raid5_error
,
9132 .hot_add_disk
= raid5_add_disk
,
9133 .hot_remove_disk
= raid5_remove_disk
,
9134 .spare_active
= raid5_spare_active
,
9135 .sync_request
= raid5_sync_request
,
9136 .resize
= raid5_resize
,
9138 .check_reshape
= raid5_check_reshape
,
9139 .start_reshape
= raid5_start_reshape
,
9140 .finish_reshape
= raid5_finish_reshape
,
9141 .prepare_suspend
= raid5_prepare_suspend
,
9142 .quiesce
= raid5_quiesce
,
9143 .takeover
= raid4_takeover
,
9144 .change_consistency_policy
= raid5_change_consistency_policy
,
9147 static int __init
raid5_init(void)
9151 raid5_wq
= alloc_workqueue("raid5wq",
9152 WQ_UNBOUND
|WQ_MEM_RECLAIM
|WQ_CPU_INTENSIVE
|WQ_SYSFS
, 0);
9156 ret
= cpuhp_setup_state_multi(CPUHP_MD_RAID5_PREPARE
,
9158 raid456_cpu_up_prepare
,
9161 destroy_workqueue(raid5_wq
);
9164 register_md_personality(&raid6_personality
);
9165 register_md_personality(&raid5_personality
);
9166 register_md_personality(&raid4_personality
);
9170 static void raid5_exit(void)
9172 unregister_md_personality(&raid6_personality
);
9173 unregister_md_personality(&raid5_personality
);
9174 unregister_md_personality(&raid4_personality
);
9175 cpuhp_remove_multi_state(CPUHP_MD_RAID5_PREPARE
);
9176 destroy_workqueue(raid5_wq
);
9179 module_init(raid5_init
);
9180 module_exit(raid5_exit
);
9181 MODULE_LICENSE("GPL");
9182 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
9183 MODULE_ALIAS("md-personality-4"); /* RAID5 */
9184 MODULE_ALIAS("md-raid5");
9185 MODULE_ALIAS("md-raid4");
9186 MODULE_ALIAS("md-level-5");
9187 MODULE_ALIAS("md-level-4");
9188 MODULE_ALIAS("md-personality-8"); /* RAID6 */
9189 MODULE_ALIAS("md-raid6");
9190 MODULE_ALIAS("md-level-6");
9192 /* This used to be two separate modules, they were: */
9193 MODULE_ALIAS("raid5");
9194 MODULE_ALIAS("raid6");